Lecture notes - Stability Theory (Math 414) Spring 2003.

Anand Pillay September 29, 2003

Introduction and preliminaries

The aim of this course and these notes is to present an exposition of the basics of stability theory, stable group theory, and geometric stability theory. I will assume knowledge of my Autumn 2002 model theory lecture notes [1]. In fact the current notes are a natural continuation of the earlier ones. The model theory lecture notes ended with a proof of Morleys Theorem. Stability theory developed historically (in the work of Shelah) as a chunk of machinery intended to help generalize Morleys theorem to a computation of the possible spectra of complete rst order theories. Here the spectrum of T is the function I (T, ), where I (T, ) is the number of models of T of cardinality . This project (at least for countable theories) was essentially completed by Shelah around 1980. In the meantime other perspectives developed, in which stability theory is seen rather as a way of classifying denable sets in a structure and describing the interaction between denable sets. Eventually this theory was seen as having a geometric meaning. This means, on the one hand, that various structural results have a geometric avour. On the other hand, it refers to the empirically discovered fact, that in numerous examples, the model-theoretic notions have actually existing mathematical meaning. This will be the perspective of these notes. We have a choice of the level of abstraction at which to operate. The theory of forking can be developed smoothly for the class of simple theories. There have been recent developments, such as rosy theories which subsume stable theories, most simple theories, and 0-minimal theories. -stable (or 1

totally transcendental) theories, on the other hand, have a somewhat more concrete avour and many natural examples t in here. An important aspect of general stability is that it is essentially a formula-by-formula theory. So we plan to nd a kind of balance between the general stable context and the -stable context

For the remainder of this introductory section we will discuss denability and imaginaries. Apart from these being important issues in their own right, this will give us a chance to give examples of the kinds of arguments (including compactness) that we will be using. Our conventions will be as follows: T will be a complete theory in a language L. T could be many-sorted, but as in the model theory notes we will usually work in the 1-sorted context. Note that there is no harm in asssuming T to have quantier elimination. Fix some big cardinal . By 4.11 of [1], T has a -saturated, strongly -homogeneous model. Let . We know that any model of T of us x such a model which we call M . So by a cardinality < is isomorphic to an elementary substructure of M model we mean (unless we say otherwise) an elementary substructure of M of cardinality < . M, N, ... will denote models. A, B, ... will usualy denote . a, b, ... will usually denote small (that is, of cardinality < ) subsets of M nite tuples of elements of M , but sometimes also small tuples. (Sometimes and tuples of elements. we will need to distinguish between elements of M In that case we write a, b, .. for elements, and a , b, .. for tuples.) , , .. will usually denote cardinals < . By a denable set we mean some subset X of n ) denotes the group of automorphisms M which is denable in M . Aut(M of M and AutA (M ) those automorphisms which x A pointwise. Note that ) acts on everything associated with M : tuples, denable sets,... Aut(M |= . For an LM = in place of M -sentence , we just write | We will be using compactness inside M in a rather specic kind of manner. Let us suppose that (y ) is a set of LA -formulas, and y a nite (or even small) tuple of variables, and that (y ) is also an LA -formula. We will write , that is, for any b (y ) |= (y ) to mean that the implication is valid in M such that |= (b, a), also |= (b). Then we have: from M Fact 1.1 (with above notation.) If (y ) |= (y ) then there is some some nite (y ) (y ) such that (y ) |= (y ). 2

Here is an application to a useful Galois-theoretic interpretation of denability over A. . Then X is A-denable Lemma 1.2 Let X be a denable set, and A M ). (or dened over A) if and only if f (X ) = X (as a set) for all f AutA (M Proof. Left to right is immediate. (Suppose (x, a) denes X where a is from n , |= (b, a) i |= (f (b), f (a)) i |= (f (b), a)) (as A. Then for any b M f (a) = a).) ). Let Right-to-left: Let (x, b) dene X (where b is some nite tuple from M p(y ) = tp(b/A). Claim I. p(y ) |= x((x, y ) (x, b)). . So tp(b /A) = tp(b/A) so there is f AutA (M Proof. Let b realize p(y ) in M such that f (b) = b . By our assumptions f (X ) = X . On the other hand ). clearly f (X ) = (x, b )(M By Fact 1.1, there is some (y ) p(y ) such that (*) |= y ( (y ) x((x, y ) (x, b))). Let (x) be the formula y ( (y ) (x, y )). Note that (x) LA . Claim II. (x) denes X . Proof. If c X , then |= (c, b), so as (y ) tp(b/A), |= (c). On the other hand if |= (c), let b be such that |= (b ) (c, b ). Then by (*) |= (c, b), so c X . By Claim II, X is A-denable. Remark 1.3 Consider the algebraic-geometric context, where T = ACFp . =K . Let V K n be an algebraic variety. Let k be a subeld of K . Write M V is said to be dened over k in the Weil-algebraic-geometric sense if the ideal [x1 , .., xn ] of polynomials vanishing on V is generated by polynomials I K with coecients from k . THEN V is dened over k in the model-theoretic sense if and only if V is dened over k p in the Weil-algebraic-geometric sense. In particular when p = 0 the two notions coincide. Here is a slight generalization of Lemma 1.2. First: n be denable. We say that X is denable almost Denition 1.4 Let X M n , such that E over A, if there is an A-denable equivalence relation E on M has only nitely many classes, and X is a union of E -classes. 3

Lemma 1.5 X is denable almost over A if and only if {f (X ) : f )} is nite. AutA (M Proof. Again left to right is immediate. (Why?) Right-to-left: Again let (x, b) dene X , and p(y ) = tp(b/A). By our as ), sumption there are b = b0 , .., bk realizing p(y ) such that for any f AutA (M ) for some i = 0, .., k . Again by saturation of M we have: f (X ) = (x, bi )(M p(y ) |= i=0,..,k x((x, y ) (x, bi )). By compactness (i.e. Fact 1.1) there is (y ) p(y ) such that |= y ( (y ) (i=0,..,k x((x, y ) (x, bi ))). Now dene E (x1 , x2 ) by the formula: (*) y ( (y ) ((x1 , y ) (x2 , y ))). Clearly E is an A-denable equivalence relation. By (*), E has nitely many classes. We leave it to you to check that X is a union of E -classes. Exercise 1.6 Show that X is denable almost over A i {f (X ) : f )} has cardinality < AutA (M . We will often identify formulas of LM with the sets they dene, so we may say (x) is over A, almost over A,.. then Note that if the denable set X is simply {a} for some a M , and 1.2 becomes: a dcl(A) i a is xed by all A-automorphisms of M 1.6 becomes a acl(A) i a has < many images under A-automorphisms . (In the latter case, if {a} is a nite union of E -classes, where E is of M A-denable and has nitely many classes, then let (x) say that {x} is an E -class, so |= (a) and (x) has only nitely many solutions.) The machinery of T eq will actually allow us to consider denable sets as elements in their own right, whereby the general and special cases in the paragraph above will be identical. T eq will be important for many other reasons. In any case in the remainder of this section we will develop T eq . The nature and status of quotient objects is rather important in many parts of mathematics, especially those with a geometric avour. For example, it is often important to know, given a manifold M and closed equivalence relation E on M , whether M/E has again (naturally) the structure of a manifold. , elements and tuples from M and From our point of view we have M denable sets of tuples, and complete types,... Suppose that E is a -denable 4

 . To what extent can we also talk about denable equivalence relation on M /E over other sets, etc. subsets of M /E , types of elements or tuples of M Namely to what extent can we include E -classes, as elements (rather than denable sets) in our whole theory. There are two approaches. The rst /E to be a set X of is somewhat informal: Dene a denable subset of M is a denable subset of M . Namely, a denable E -classes such that X M set of E -classes is something coming from an E -invariant denable subset of . Dene tp((b/E )/A) to be the set of A-denable subsets of M /E which M contain b/E . Then show that all our tools, such as compactness hold for the generalized denable sets, types, etc... In fact this informal approach is how we deal with hyperimaginaries. The more formal approach is to contruct a new complete many-sorted eq (whose theory is a rst order many sorted theory T eq ) in structure M which all these quotient objects are by denition elements. Let us rst dene a many-sorted language Leq (which actually depends on T as well as L, in spite of the notation). For each L-formula (x1 , .., xn , y1 , .., yn ) such that T says that denes an equivalence relation on n-tuples, let S be a sort symbol. (Thinking semantically we sometimes write SE in place of S ). So among the sorts is S= . Let us also introduce, for each such (x1 , .., xn , y1 , .., yn ) a new function symbol f whose domain sort is n-tuples of sort S= and whose range sort is S . Finally for every m-place relation symbol R of L, R will still exist in Leq but as a relation on m-tuples of sort S= . Likewise for old function symbols of L. Note that any L-sentence can be identied with an Leq -sentence: just take all variables to be of sort S= . (Note that by denition of the many-sorted logic, if x is a variable of a given sort S then x(...) is interpreted as for all x of sort S .) We dene T eq to be the Leq -theory axiomatized by: (i) T , (ii) for each as above, the sentence x1 , .., xn , y1 , .., yn (of sort S= ) ((x1 , .., xn , y1 , .., yn ) f (x1 , .., xn ) = f (y1 , .., yn )). (iii) for each as above, the sentence expressing that f is surjective (as a map from (S= )n to S ). Finally for M a model of T , M eq will denote the Leq -structure, such that S= (M eq ) is the universe of M , S (M eq ) = M n /E (where E is the equivalence relation on M n dened by ), f (M eq )(a1 , ., an ) = (a1 , .., an )/E (for a1 , .., an M n ), and such that the old relation and function symbols of L are interpreted on S= (M eq ) as they were in the L-structure M . 5

We may often notationally identify S= (M eq ) with M , where hopefully there is no confusion. So we have our original L-structure, living, equipped with all its original L-structure as a sort in M eq . The rst question a model-theorist asks in this kind of situation is whether M gets any more induced structure this way. We will point out below, among other things, that M does NOT get any new structure. Lemma 1.7 (i) The models of T eq are precisely the structures M eq for M a model of T . (ii) If M, N are models of T then any isomorphism between M and N extends to a (unique) isomorphism between M eq and N eq . (iii) T eq is a complete theory. (iv) If M, N are models of T , a is a tuple (maybe innite) from M and ba tuple from N , and tpM ( a) = tpN ( a) then tpM eq ( a) = tpN eq (b). Proof. (i) and (ii) are obvious. For (iii) to make life a bit easy let us assume GCH so there are arbitrarily large cardinals in which any given theory has a saturated model. Let M eq , N eq be saturated models of T eq of cardinality some |T |. So the same is true of M, N . So as T is complete M and N are isomorphic. Hence by (ii), M eq and N eq are isomorphic, so elementarily equivalent. (iv) Same proof as (iii). Fix n. Let us consider two Stone spaces. First W1 = Sn (T ), the space of complete n-types of T . Secondly W2 = S(S= )n (T eq ), the space of complete types of n-tuples from sort S= in models of T eq . (Sorry for the double meaning of S here.) We have a natural restriction map : W2 W1 which is continuous. On the other hand by Lemma 1.7 (iv) and a saturation argument, is a a bijection. Thus: Lemma 1.8 is a homeomorphism. We conclude: Proposition 1.9 Let (x1 , .., xk ) be an Leq -formula, where xi is of sort SEi say. Then there is an L-formula ( y1 , .., y k ) such that T eq |= y 1 ..y k ( ( y1 , .., y k ) (fE1 ( y1 ), .., fEk ( yk )).

Proof. Let n be the length of the tuple y = ( y1 , .., y k ). By the previous 1 lemma {p( y ) Sn (T ) : (fE1 ( y1 ), .., fEk ( yk )) (p)} is clopen, which is enough. Corollary 1.10 (i) If M |= T is an elementary substructure of N then M eq is an elementary substructure of N eq . eq is also (ii) M -saturated and -strongly homogeneous. The proof is left as an exercise. correponds to an element of M eq : Suppose X M n Any denable set in M is dened by (x, b). Let E (y1 , y2 ) be the L-formula x((x, y1 ) (x, y2 )). eq . Let us note a couple of things: Then b/E is a element of sort SE in M (or equivalently of M eq ), (X ) = X Firstly, for any automorphism of M i (b/E ) = (b/E ). eq . This can either Secondly, X can be dened over b/E in the structure M eq , or directly by be seen using Lemma 1.2 in the many sorted structure M considering the formula y (fE (y ) = b/E (x, y )). Thirdly, let (x, (b/E )) be the Leq -formula with parameter b/E which denes X . Then b/E is the unique element z of sort SE such that (x, z ) denes X . Sometimes we say that the imaginary b/E above codes the denable set X . In fact for an arbitrary -denable equivalence relation E say (on say) and c M n , c/E codes the set dened by E (x, c). n-tuples from M eq as adjoining codes for all denable sets in M . So we can also think of M Actually the code for X is not unique, as it depends on the shape of the formula used to dene X . But it is unique up to interdenability. Let us look back at Lemmas 1.2 and 1.5 in this light. Let X be a denable , let e M eq be a code for X . Then X is denable over A in M i set in M eq . Likewise, X is denable almost over A in M i e acl(A) e dcl(A) in M eq . Let us make this precise: in M n be denable, and let A M . Then X is denLemma 1.11 Let X M eq . able almost over A if and only if X is acleq (A)-denable in M eq . So Proof. Let e be the code for X . So X is denable over e in M eq if X is almost over A, then e acl(A) and X is acl (A)-denable in eq . Conversely, suppose X is acleq (A)-denable. As e is a code of X , M 7

e dcleq (acleq (A)) = acleq (A). So e has nitely many images under A eq , as does therefore X too. So by 1.5, X is almost over automorphisms of M A. . a and b are said to have the Denition 1.12 Let a, b be n-tuples from M same strong type over A (we write stp(a/A) = stp(b/A)) if whenever E is an n with only nitely many classes then A-denable equivalence relation on M E (a, b). Note that stp(a/A) = stp(b/A) implies tp(a/A) = tp(b/A). With the above notation: Lemma 1.13 The following are equivalent: (i) stp(a/A) = stp(b/A), (ii) a and b satisfy the same formulas which are almost over A. eq ). (iii) tp(a/acleq (A)) = tp(b/acleq (A)) (in M Proof. (i) and (ii) are trivially equivalent. The equivalence of (ii) and (iii) follows immediately from Lemma 1.11. Now it may happen that we can already nd codes for denable sets in M itself. When this happens we say that T admits elimination of imaginaries. Denition 1.14 T has elimination of imaginaries if for any denable X n , and code e M eq for X , there is some nite tuple c from M such M eq ). (Alternatively, without mentioning that e dcl(c) and c dcl(e) (in M eq such that an automorphism of M M we simply require c to exist in M xes X setwise i it xes the tuple c.) Exercise 1.15 The following are equivalent: (i) T has elimination of imaginaries. (ii) If M |= T and e M eq then there is a nite tuple c from M such that dcl(c) = dcl(e) in the sense of M eq . (iii) Let M |= T , (x, y ) L and b in M . Then there is a formula (y ) tp(b), and a partial -denble function f with range contained in some M k such that f is dened on (M ) and for all b1 , b2 satisfying (y ), f (b1 ) = f (b2 ) i (x, b1 )(M ) = (x, b2 )(M ). (iv) Let again M |= T and (x, y ) L where y is an n-tuple. Then M n can 8

be partitioned into nitely many -denable sets Y1 .. Yr say and there are -denable functions fi : Yi M ki for i = 1, .., r such that for each i and b1 , b2 Yi , fi (b1 ) = fi (b2 ) i (x, b1 )(M ) = (x, b2 )(M ). (v) For any denable set X in a model M of T there is an L-formula (x, y ) and b such that b is the unique tuple from M such that (x, b)(M ) = X . Example 1.16 Let T be the theory of an equivalence relation with exactly 2 classes both innite. Then T does not eliminate imaginaries. In the above we dened T eq for T 1-sorted, but exactly the same thing can be done for T many-sorted. Similarly, the notion of eliminating imaginaries makes sense for T many-sorted. Lemma 1.17 T eq eliminates imaginaries. Proof. Let E be a -denable (in Leq ) equivalence relation on some sort SE say in a model M eq of T eq . By Proposition 1.9, there is an L-formula (y1 , y2 ) (where yi are of appropriate length) such that for any a1 , a2 , M |= (a1 , a2 ) i M eq |= E (fE (a1 ), fE (a2 )). So clearly denes an equivalence relation, and we have a -denable bijection between the sorts SE /E and S . eq , stp(a/A) = stp(a/A) i Note that it follows, for example, that in M tp(a/acl(A)) = tp(b/acl(A)), where a, b are tuple from any sorts and A is a set of elements from arbitrary sorts. Often we will work in T eq . Many algebraic theories/structures such as algebraically closed elds, real closed elds, dierentially closed elds, eliminate imaginaries. But some very basic theories such as the theory of an innite set with no structure do not: the point being only that nite sets (which are denable) need not be coded. Denition 1.18 T has weak (geometric) elimination of imaginaries if for some model M of T and e M eq there is a nite tuple c from M such that e dcl(c) (e acl(c)) and c acl(e) (in M eq ). Exercise 1.19 Show that the theory of an innite set (in the language with only equality) has weak elimination of imaginaries, but does not have (full) elimination of imaginaries. 9

Hint. Unravelling things, we have to show that for any denable set X in a model, there is a nite tuple c such that c is xed by any automorphism xing the set X , and that X has only nitely many images under the group of automorphisms xing c. Use quantier-elimination.

Stability, totally transcendental theories, and forking

T is as in section 1, one-sorted for convenience. We work in the big model . At some point we will pass to T eq . In [1] we dened stability in terms M of counting types. We will take a dierent approach here but will soon show the various denitions to be equivalent. We will use the -stable theories dened in [1] as providing somewhat more concrete examples/applications of the machinery developed here. stability was dened only for countable theories. As we are no longer assuming T to be countable, the appropriate notion is that of a totally trascendental (or t.t) theory. Remember that we dened the Morley rank of an arbitrary formula (with parameters) and this still makes sense for uncountable T . If (x) is a formula (with parameters, where x is an n-tuple of variables), and ) the subset of M n dened by we will use RM (X ) and RM () X = (M interchangeably. Denition 2.1 T is t.t if for every denable set X , RM (X ) < . (Here n for some n.) X is a denable subset of M Exercise 2.2 (i) If RM (x = x) < (x a single variable), then T is t.t. (ii) If T is t.t then T eq is t.t.. Namely for every denable subset X of a nite eq ). product of sorts, RM (X ) < (computed in M Recall also that RM (tp(c/A)) is by denition min{RM ((x)) : (x) tp(c/A))}, and if this is ordinal valued and witnessed by (x) then dM (tp(c/A)) = dM ((x)). We will be dening a notion of independence: for c, d nite tuples and A a set of parameters, we will make sense out of the expression c is independent from d over A (at least for T stable). In the case where T is t.t, this will reduce to RM (tp(c/A, d)) = RM (tp(c/A)). In the even more 10

special case where T is the theory of algebraically closed elds and A = k is a subeld it will reduce to tr.deg.(k (c, d)/k (d)) = tr.deg.(k (c)/k ), that is the elds k (c) and k (d) are algebraically disjoint over k . In fact in this last example we will see that RM (tp(c/k )) = tr.deg.(k (c)/k ). The independence notion will come from Shelahs theory of forking. This is also valid for simple theories, but in the stable case there are some special features such as (i) a theory of multiplicity (like Morley degree) and (ii) the theory operates and can be developed on a formula-by-formula basis. Denition 2.3 Let (x, y ) be a formula, possibly with parameters). We say that (x, y ) is stable if there do not exist ai , bi for i < such that for all i, j < , |= (ai , bj ) i i < j . Remark 2.4 (i) The stability of (x, y ) depends on the choice of a division of the tuple of variables in into two subtuples. (ii) If (x, (y, z )) is stable then so is (x, y, c) for any tuple c of parameters. (iii) If 1 (x, y ), 2 (x, z ) are stable, then so are (x, y ), (1 2 )(x, (y, z )), (1 2 )(x, (y, z )), and 1 (y, x) = (x, y ). (iv) Suppose (x, y ) is stable. Then there is some N < such that there do not exist ai , bi for i N such that |= (ai , bj ) i i < j . (v) Suppose that (x, y ) is unstable. Let ((x1 , y1 ), (x2 , y2 )) be the formula (x1 , y2 ). Then there are (ci : i < ) such that |= (ci , cj ) i i < j . Proof. (ii) is obvious. (iii) is a exercise, and (iv) follows from compactness. (v) Let ai , bi for i < witness the instability of . Put ci = (ai , bi ). The following begins to tie up stable formulas with the number of types. Lemma 2.5 (i) Suppose that some L-formula (x, y ) is unstable. Then for any |T |, there is a model M of T such that there are > many complete types (in variable x) over M . In fact we can nd such a set P of types which are distinguished by formulas of the form (x, b) for b M (ii) If T is t.t. then every L-formula is stable. Proof. Let be the least cardinal such that 2 > . So mu . Let 2 be the set of functions from to 2, and order it by: f < g if there is < such that f | = g | and f () = 0 and g () = 1. By compactness there are af , bf for f 2 such that |= (af , bg ) i f < g . Let X 2 the set of eventually 11

constant functions. Then |X | , and for each f1 = f2 2 there is g X such that |= ( (af1 , bg ) (af2 , bg )). Let M be a model of cardinality containing {bg : g X }, and we see that the set {tp(af /M ) : f 2} satises our requirements. (ii) If T is t.t. then as in the proof of 5.19 of [1] there are at most many complete types over any model of cardinality |T |, so we can use (i). The next lemma is crucial: a tuple Lemma 2.6 Let (x, y ) be stable. Let M be a model, and a M j of length l(x). Then there is a nite set {ai }i,j<N of tuples contained in M such that (*) for any b M , |= (a, b) i |= j (i (aj i , b)). Moreover (i) N is determined only by (not by the choice of a or M ), (ii) if (x, c) is any formula over M satised by a we can choose the aj i to satisfy (x, c), and (iii) if M is -saturated and C is a subset of M of cardinality < , we may choose the aj i to realize tp(a/C ). Proof. The proof will be in two steps: First let N1 be as given by 2.4 (iv). Claim. For any formula (x) tp(a/M ) there are a1 , .., an M (all satisying (x)) for some n N1 such that for any b M , if M |= i=1,..,n (ai , b) then |= (a, b). Proof. We start to build a1 , a2 , ... in M satisfying (x) and b1 , b2 , .. in M such that |= (a, bi ), and such that |= (ai , bj ) i i j . Suppose we have already found such a1 , .., an , b1 , .., bn . Then the formula (x) i=1,..,n (x, bi ) is over M , and true of a, so realized in M by an+1 say. Now either (a1 , .., an+1 ) satises the claim, and we stop, or there is b M such that |= i=1,..,n+1 (ai , b) but |= (a, b). Let bn+1 M be such a b. So the construction continues. So we have to stop for some n N1 . We now pass to Step 2. Let ((x1 .., xN1 ), y ) be the formula (x1 , y ) ... (xN1 , y ). By Remark 2.4, is stable (with the given division of its variables). Let N2 be as given by 2.4 for . Let x denote (x1 , .., xN1 ). We construct 1 2 a ,a , .. in M satisfying the claim above, and b1 , b2 , ... in M such that (i) |= ( ai , bj ) i i > j , 12

(ii) |= (a, bi ) for all i. Suppose we already have constructed a i , bi for i = 1, .., n. By the claim there is a n+1 in M such that |= ( an+1 , bi ) for i = 1, .., n, and such that for any n+1 b M , if |= ( a , b) then |= (a, b). So if i=1,.,n+1 ( ai , y ) does not satisfy the requirements of the lemma then there is bn+1 M such that |= (a, bn+1 ) but |= ( ai , bn+1 ) for i = 1, .., n + 1. So the construction continues. It follows that we must stop at some n N2 . This proves the existence of the { aj i }. (i), (ii) and (iii) follow fron the construction and the claim. Exercise 2.7 Let (x, y ) L and let M be a model such that there is no innite ordered set (I, <) and ai , bi in M for i I such that |= (ai , bj ) there are a1 , .., an M such that for i i < j . Prove that for any a M any b M whether or not (a, b) holds depends only on which (ai , b) hold. Namely there is a formula (y ) equivalent to some Boolean combination of the (ai , y ) such that for any b M , |= (a, b) i M |= (b). Hint. Assume not. Then construct ai , bi , ci in M for i < such that (1) for each i, |= (a, bi ) (a, ci ). (2) for each i j , |= (ai , bj ) (a, bj ) and |= (ai , cj ) (a, cj ), and (3) for each i < j , |= (ai , bj ) (ai , cj ). By thinning out the sequence and applying Ramseys theorem get a contradiction to our assumptions. Denition 2.8 (i) Let (x, y ) L. By a complete -type over a model M , we mean a maximal consistent set of formulas of the form (x, b) or (x, b) where b M . S (M ) denotes the set of such types when there is no ambiguity. (ii) The complete -type p(x) S (M ) is said to be denable if there is an LM -formula (y ) such that (M ) = {b M : (x, b) p(x)}. Note that (y ) is unique up to equivalence. If A M we say that p(x) S (M ) is denable over A if (y ) is over A (that is equivalent to an LA -formula) and denable almost over A, if (y ) is almost over A. (iii) The complete type p(x) S (M ) is denable if for every (x, y ) L, p| is denable, and denable (almost) over A if each p| is. (iv) Suppose p(x) S (M ) is denable, and for each (x, y ) L let be as in (ii). We call the map a dening scheme for p(x). Corollary 2.9 Let (x, y ) L. The following are equivalent: (i) is stable, 13

(ii) Every complete -type over a model is denable, (iii) For each |T |, and model M of cardinality , |S (M )| . Proof. (i) implies (ii) is given by 2.6. (But note that we showed there that the dening formula can be chosen to be of a very special form.) (ii) implies (iii): Given M |= T , any p(x) S (M ) is clearly determined by a dening formula (y ) for p (as in the denition). But there are at most |M | + |T |-many such formulas. (iii) implies (i): By Lemma 2.5 (i). Corollary 2.10 The following are equivalent: (i) every formula (x, y ) (any variables x, y ) is stable, (ii) every complete type over a model is denable, (iii) For any |T | and M |= T of cardinality , the set of complete types over M has cardinality at most |T | , (iv) There is some |T | such that for every model M of cardinality , S (M ) has cardinality . (v) every formula (x, y ) with x a single variable (or a single variable from some sort) is stable. Proof. This is almost immediate from Corollary 2.9. For (ii) implies (iii), note that the number of possible dening schema for types over M is |M ||T | . Denition 2.11 We will say that T is stable if it satises the equivalent conditions of Corollary 2.10. Exercise 2.12 If T is stable, then so is T eq . We will now specialize somewhat to t.t theories and return to the general theory later. First some notation: let p(x) S (M ) be a denable type, and let d denote the corresponding dening schema, that is d( (x, y )) = (y ) as in 2.8 (iv). Let N be a model containing M . By d(N ) we mean { (x, b): for (x, y ) L and b N such that |= d( )(b)}., Exercise 2.13 With this notation, d(N ) is a complete type over N , extending p(x). 14

Lemma 2.14 Let (x) LA be a formula such that RM ((x)) = < and : RM ((x) (x, b)) = dM ((x)) = 1. Let (x, y ) L be stable. Then {b M } is denable, over A. Proof. We may assume A to be nite. Let M be an -saturated model containing A. Let p(x) S (M ) be such that (x) p(x) and RM (p(x)) = . (Why does such p exist?) By 2.10, p(x)| is denable, by a formula (y ) LM say. But clearly for b M , (x, b) p(x) i RM ((x) (x, b)) = . Thus for b M , M |= (b) i RM ((x) (x, b)) = . As M is -saturated, . (Why?) Also (x) is equivalent to a the equivalence is valid for all b M formula over A, by Lemma 1.2. Corollary 2.15 Let (x) be a formula of Morley rank < and Morley : degree d possibly greater than 1. Let (x, y ) L be stable. Then {b M (RM, dM )((x) (x, b)) = (, d) is denable over A. Proof. Let 1 (x), .., d (x) be formulas of Morley rank and Morley degree 1 whose disjuction is equivalent to (x). Then (RM, dM )((x) (x, b)) = (, d) if and only if RM (i (x) (x, b)) = for i = 1, .., d. So we can use lemma 2.14 (and Lemma 1.2). Proposition 2.16 Suppose T is t.t, and p(x) S (M ) (M a model). Then p has Morley degree 1. Proof. Suppose not. So let (x) p(x) be such that (RM, dM )(p(x)) = (RM, dM )((x)) = (, d) with d > 1. For each (x, y ) L let (y ) be the formula over M given by Corollary 2.15. Then one can see that actually the which schema d( ) = (y) denes p(x). Let (x, d) be a formula over M implies (x), and has Morley rank and Morley degree 1. Let N be a model containing M and d. Let q (x) = d(N ), which is by 2.13 a complete type over N extending p(x). Now both (x) (x, d) and (x) (x, d) have (RM, dM ) < (, d), so |= (d) and |= (d), a contradiction (why?). Exercise 2.17 Let T be t.t and p(x) S (M ). Then there is a nite subset A of M such that p(x) is denable over A. Corollary 2.18 Let T be t.t., let M < N be models, p(x) S (M ) and p(x) q (x) S (N ). Then q (x) is denable over M if and only if RM (q (x)) = RM (p(x)). 15

Proof. Suppose q is denable over M , with dening schema d. So clearly d(M ) = p(x). Let p(x) have Morley rank and degree 1 (by 2.16), and let this be witnessed by (x) p(x). So for (x, y ) L, d( )(y ) is equivalent to the formula RM ((x) (x, y ) = which is given by 2.14 (in M so ). As q = d(N ) too we see that for each formula (x, b) q (x), also in M RM ((x) (x, b)) = , hence RM (q ) = . Conversely, suppose RM (p) = RM (q ) = . Let (by 2.16) (x) p(x) have Morley rank and Morley degree 1. For each (x, y ) let (y ) be RM ((x) (x, y )) = . Then clearly ( ) is a dening schema (over M ) for q . Exercise 2.19 Assume that p(x) S (A) and RM (p) = < . Then for any B A there is q (x) S (B ) such that p(x) q (x) and RM (q ) = . Hint. By Exercise 5.18 of [1], p(x) {(x) : (x) LB , RM ((x)) < } is consistent. Lemma 2.20 Suppose T is t.t. (i) Suppose A M , p(x) S (A) and p(x) q (x) S (M ) with RM (p) = RM (q ) = . Then q is denable almost over A. (ii) Suppose p(x) S (A) and A M . Then there is some q (x) S (M ) with p(x) q (x) such that q is denable almost over A. Proof. (i) We may assume that M is (|T | + |A|)+ -saturated and strongly homogeneous. Note that if f is an A-automorphism of M , then f (q ) is an extension of p(x) of Morley rank hence there are only nitely many possibilities for f (q ). (Why?) In particular, for each (x, y ) L, the -denition, (y ) say, of q has only nitely many images under A-automorphisms of M , hence is almost over A. (ii) Suppose p(x) has Morley rank . By Exercise 2.19, p(x) has an extension in S (M ) of RM . Now use part (i). Lemma 2.21 Let T be stable. Let p(x), q (y ) be complete types over M which are denable over A. Let (x, y ) L, and write it also as (y, x). Let (y ) be the (x, y )-denition of p(x), and (x) the (y, x)-denition of q (y ). (So (y ), (x) can be assumed to be LA -formulas.) Then (y ) q (y )|A if and only if (x) p(x)|A. 16

Proof. We may assume M to be saturated enough. Suppose for a contradiction that (y ) q (y ) but (x) p(x). Construct ai , bi in M for i < such that a1 realizes p(x)|A, b1 realizes q (y )|(A, a1 ), and in general an+1 realizes p(x)|(A, b1 , .., bn ) and bn+1 realizes q (y )|(A, a1 , .., an+1 ). Note that |= (ai ) (bi ) for all i, so clearly |= (ai , bj ) i i > j , contradicting stability of (x, y ). We will now use T eq . Corollary 2.22 Assume T to be t.t.. Let p1 (x), p2 (x) S (M ) be both denable over A M . Suppose that p1 |acleq (A) = p2 |acleq (A). Then p1 = p2 . Proof. Let (x, y ) L and b M . We want to show that (x, b) p1 (x) i (x, b) p2 (x). Namely that |= 1 (b) i |= 2 (b) where i (y ) LA is the -denition of pi (x). By 2.20 (ii) tp(b/A) has an extension q (y ) S (M ) which is denable almost over A. In particular if (x) is the (y, x)-denition of q (y ) (where (y, x) = (x, y )) then (x) is over acleq (A). By Lemma 2.21 eq ), (x) pi |acleq (A) i i (y ) q (y ), for i = 1, 2. (which is also valid in M So by our assumptions, |= 1 (b) 2 (b). Here is a restatement: Corollary 2.23 Assume T to be t.t. Let A = acleq (A) and p(x) S (A). Then for any model M containing A there is a unique extension q (x) S (M ) of p(x) such that q (x) is denable over A. Note in particular that, in the context of 2.23, for any B containing A there is a unique extension q (x) S (B ) of p(x) such that some extension of q (x) to a model is denable over A. So (again in the context of 2.23) we will denote this extension by p(x)|B (hopefully without ambiguity). Moreover there is a dening schema d over A such that for any B A, d(B ) = p|B . Denition 2.24 (T stable). (i) Let p(x) S (A), A B and p(x) q (x). We say that q (x) is a nonforking extension of p(x) (or q (x) does not fork over A) if q (x) has some extension q (x) to a complete type over a model, such that q (x) is denable almost over A. (Here x is a nite tuple of variables.) (ii) Suppose A B , and a a nite tuple. We will say that a is independent 17

from B over A if tp(a/B ) does not fork over A. (iii) If a is a possibly innite tuple, we say that a is independent from B over A if a is independent from B over A for all nite subtuples a of a, (iv) If a and b are arbitrary tuples, we will say that a is independent from b over A, if a is independent from A b over A in the sense of (iii). Exercise 2.25 The notion of nonforking in (i) above is invariant under adding names (constants) for elements of A. Proposition 2.26 Suppose T to be t.t.. Let p(x) S (A), A B and p(x) q (x) S (B ). Then q (x) is a nonforking extension of p(x) if and only if RM (p) = RM (q ). Proof. The denitions, as well as Exercise 2.19 allow us to assume that B is a model M . The right to left direction is given by Lemma 2.20 (i). Left to right: Let RM (p(x)) = . Let p1 (x) = q |acleq (A). The rst observation is that RM (p(x)) = RM (p1 (x)). This is because, for any formula (x) p1 , some disjunction of nitely may A-conjugates of is over A, hence in p(x). Now, by 2.19 and 2.20 (i), there is an extension q1 (x) S (M ) of p1 (x) which is denable over acleq (A), and such that RM (q1 ) = RM (p1 ) = RM (p). By 2.23, q1 = q . Remark 2.27 Let us make explicit something observed in the proof above (for T t.t): for any a, A, a is independent from acleq (A) over A. Exercise 2.28 Let p(x) S (A). Let p1 (x), p2 (x) S (acleq (A)) be extensions of p(x). Show that there is an elementary f : acleq (A) acleq (A) such that f |A = id and f (p1 (x)) = p2 (x). Proposition 2.29 (Assume T to be t.t.) Let a, b denote nite tuples, and x a nite tuple of variables. (i) (transitivity) Suppose that A B C . Then a is independent from C over A i a is independent from C over B and from B over A. (ii) (symmetry) a is independent from b over A i b is independent from a over A. (iii) for any a and A there is a nite A0 A such that a is independent from A over A0 , (iv) For any type p(x) S (A) the number of nonforking extensions of p(x) 18

over some (any) model is precisely the Morley degree of p. (v) (existence) For any p(x) S (A) and B A, p(x) has a nonforking extension q (x) S (B ). (vi) (nite character) Suppose that A B and p(x) S (B ) forks over A. Then there is a nite tuple b from B such that p(A)|(A b) forks over A. Proof. (i) is immediate from 2.26. (ii) We may assume A = acleq (A) (by 2.27). Let p(x) = tp(a/A) and q (y ) = tp(b/A). By Exercise 2.25 we may assume that the elements of A are named by constants. Let d1 be the dening schema corresponding to p given after 2.23, and d2 the one for q . So a is independent from b over A i a realizes d1 (A b) and b is independent from a over A if b realizes d2 (A a). Now suppose b to be independent of A over a. Let (x, y ) L, and suppose (x, b) d1 (A, b). We want to show that |= (a, b). Now |= d1 ( )(b). By 2.21, |= d2 ( )(a), where (y, x) = (x, y ). Thus, (y, a) d2 (A, a), whereby |= (b, a), namely |= (a, b). (iii) Choose a formula (x) tp(a/A) with Morley rank = RM (p). So (x) has parameters from a nite subset A0 of A. Then (x) p|A0 , so RM (p|A0 ) = . Now use 2.26. (iv) Let p(x) S (A), and suppose (RM, dM )(p) = (, d). Let M A be any model, and let M > M be suitably saturated. It is easy to show that p(x) has precisely d extensions p1 , .., pd to complete types over M of Morley rank (and degree 1). For each i = 1, .., d, pi |M has Morley rank and degree 1 (by 2.16). (v) By 2.20 (ii), or by 2.19 and 2.26. (vi) If p(x) forks over A then by Proposition 2.6, RM (p) < RM (p|A). Suppose RM (p) = . Choose a formula in p of Morley rank and let b be the tuple of parameters in that formula. Remark 2.30 In the light of Denition 2.24, (i), (ii), and (vi) of 2.29 are also valid for a, b possibly innite tuples. In fact (v) (existence) is also true for complete types of innite tuples. Proof. Let us just consider existence. Fix an innite tuple a = (a1 , a2 , ...) say, and a set A. Let M be a model containing A. Let pn (x1 , .., xn ) = tp(a1 , .., an /acleq (A)). Let qn (x1 , .., xn ) S (M ) be the unique nonforking extension of pn to M (by 2.23). Note that qn (x1 , .., xn ) qn+1 (x1 , .., xn+1 ) 19

for all n. (Why?) Let q (x1 , x2 , .....) be the union, a complete type over M . If b = (b1 , b2 , ...) ralizes q , then b is independent from M over A and tp(b/A) = tp(a/A). Corollary 2.31 (T t.t.) Assume A C . Then tp(a, b/C ) does not fork over A if and only if tp(b/C ) does not fork over A and tp(a/Cb) does not fork over Ab. Proof. Enumerate C as c. Assume that (a, b) is independent from c over A. By Remark 2.30 (symmetry), c is independent from (a,b) over A. By transitivity, c is independent from b over A and clearly c is independent from a over A b. Apply symmetry again to conclude that b is independent from c over A and a is independent from cb over Ab. The converse goes the same way. Denition 2.32 (T t.t.). Let p(x) S (A). We say that p(x) is stationary if it has exactly one nonforking extension over any B A. Of course this denition makes sense for an arbitrary theory, but for T t.t., stationarity of p(x) S (A), is equivalent, by 2.29 (iv) to p having Morley degree 1. Moreover, any complete type over an eq -algebraically closed set is stationary (by 2.23). Proposition 2.33 (Conjugacy of nonforking extensions, nite equivalence relation theorem.) Assume T to be t.t. (i) Let p(x) S (A). Let M A be (|T | + |A|)+ -saturated and strongly homogeneous. Let p1 (x), p2 (x) S (M ) be nonforking extensions of p(x). Then there is an A-automorphism f of M such that f (p1 ) = p2 . (ii) Let (x) be a formula over A of Morley rank and Morley degree d. Then there is a formula (x) over A, which implies (x) and such that RM ((x) (x)) < , and there is an A-denable equivalence relation E (x1 , x2 ) on ) with precisely d classes, each of Morley rank . (M Proof. (i) First note that if q (x) S (M ) is a nonforking extension of p(x) then q (x) is determined by q |acleq (A) (by 2.23). So now, suppose q1 (x), q2 (x) S (M ) are nonforking extensions of p(x). By Exercise 2.28, let f be an elementary permutation of acleq (A) which xes A pointwise and takes q1 |acleq (A) to q2 |acleq (A). f then extends to an automorphism f of 20

M , and f (q1 ) = q2 . (ii) We begin with Claim. We can nd pairwise inconsistent formulas 1 (x), .., d (x) over acleq (A), each of which implies (x) and each of which has Morley rank and Morley degree 1. Proof. Let 1 (x), .., d (x) be pairwise inconsistent formulas of Morley rank and degree 1 over some model M A, each of which implies (x). Let qi (x) S (M ) contain i (x) and have Morley rank . By 2.26, qi (x) does not fork over A ((x) qi (x), so RM (qi |A) = ). Let pi (x) = qi (x)|acleq (A). Then pi (x) is stationary, (so of Morley degree 1) and i = j implies pi (x) = pj (x). So we may nd formulas i (x) in pi (x) of Morley rank and degree 1. We may assume that the i (x) are pairwise inconsistent. Let i be the (nite) set of images of i (x) under A-automorphisms (i (x) is almost over A). Let be i i . So is a nite set of formulas, and moreover is A-invariant (in the obvious sense). Let E (x1 , x2 ) be: {(x1 ) (x2 ) : (x) }. Then E is denable, A-invariant and so A-denable. Moreover ). As each formula in E has nitely many classes. Let us restrict E to (M has Morley rank and degree 1, it is clear that each E -class either has Morley rank and degree 1 or Morley rank < . Let X be the union of the classes of Morley rank . So X is denable and A-invariant, so A-denable, by (x) say. Then (x) and E satisfy the required conditions. We will discuss a few more issues in the context of t.t theories (such as canonical bases, Morley sequences, dividing, existence of saturated models), and then discuss how the whole theory generalizes to the stable case. Denition 2.34 (T stable.) Let p(x) be a global type, that is a complete . Let c be a small tuple from M eq . We say that c is a canonical type over M base of p if for any automorphism f of M , f (p) = p i f (c) = c. Usually canonical bases are thought of as sets. Here we think of them as (possibly innite) tuples, although this also causes some clumsy notation. In the case of algebraically closed elds, say, any stationary type is the generic type of some irreducible variety, and the canonical base of the type identies with the smallest eld of denition of the variety. Canonical bases play a crucial role in both general stability theory and geometric stability theory. In eq ), and given a the t.t case canonical bases will exist as single elements (in M 21

stationary type p, if c is the canonical base of the global nonforking extenion of p, RM (tp(c)) will measure the size of the set of the set of conjugates ). Algebraically closed elds of p (images of p under automorphisms of M are an interesting example. Consider the family of algebraic curves over C given by y = an xn + an1 xn1 + .. + a0 (as the ai vary). This is clearly an n + 1-dimensional family of curves. This corresponds to the fact that for generic independent a0 , .., an , the tuple (a0 , ..., an ) is a canonical base for the generic type of the curve, and RM (tp(a0 , .., an )/)) = n + 1. Proposition 2.35 (T stable) Any global type p(x) has a canonical base, which is unique up to interdenability. Namely if c1 , c2 are canonical bases of p then c1 dcl(c2 ) and c2 dcl(c1 ). In particular, p has a unique eq ), up to enumeration, which we refer denably closed canonical base (in M to as Cb(p). Proof. For each (x, y ) L, let (y ) LM (by 2.10), be a -denition for p eq let c be a code in M for (y ), and let c = (c ) . It is clear that c works. If d is another canonical base for p, then by denition, an automorphism xes c i it xes d. Hence c and d are interdenable. of M Denition 2.36 (T t.t.) Let p(x) S (A) be stationary. We dene Cb(p) to be Cb(p(x)) where p(x) is the unique global nonforking extension of p. Remark 2.37 (T t.t.) So by the proof of 2.35, if p(x) S (A) is stationary, then Cb(p) is the eq -denable closure of the set of codes of -denitions of p (the global nonforking extension of p) as (x, y ) ranges over L. Let me reiterate a remark made earlier: Any complete type p(x) over a set A has a unique extension to a complete type over dcleq (A). In particular it makes sense to talk about the restrition of p(x) to C for C dcleq (A). Lemma 2.38 (Assume T to be t.t.) (i) Let p(x) S (A) be stationary. Then Cb(p) is the smallest eq -denably closed subset A0 of dcleq (A) such that p(x) does not fork over A0 and p|A0 is stationary. (ii) Given a nite tuple a and A B , tp(a/B ) does not fork over A if Cb(stp(a/B )) is contained in acleq (A). (iii) If p(x) S (A) is stationary, then Cb(p) is the eq -denable closure of a 22

 eq ). single element (of M (iv) (normalization) Let (x) be a formula of Morley rank and degree 1. Then there is another formula (x) of Morley rank and degree 1, such that (RM, dM )((x) (x)) = (, 1) and for any conjugate (x) of (x) under ), either |= x( (x) (x)), or RM ( (x) (x)) < . Aut(M xes p Proof. (i) Let c be Cb(p). Note that any A-automorphism of M (the global nonforking extension of p) so xes c pointwise, whereby c dcleq (A). Suppose that A0 dcleq (A), and p does not fork over A0 and p|A0 is stationary. Then clearly the global nonforking extension of p|A0 is the same as the global nonforking extension of p. Therefore by the rst part of the proof applied to p|A0 we see that c dcleq (A0 ). Finally we need to show that p(x) does not fork over c and p(x)|c is stationary. In fact by denition (Denition 2.24 (i)) p(x) does not fork over c. To show that p(x)|c is stationary, we show that p is its unique global nonforking extension. Let q be another global nonforking extension of p|c. By 2.33 (i) there is a c taking p to q. But p is xed by all c-automorphisms of automorphism of M . So q = p. M (ii) Remember that stp(a/B ) denotes tp(a/acleq (B ), a stationary type. Let c = Cb(stp(a/B )). Let p be the global nonforking extension of stp(a/B ). Suppose that c acleq (A) where A B . Then clearly p is denable almost over A, so does not fork over A. By transitivity p|B does not fork overA. That, is tp(a/B ) does not fork over A. Conversely assume tp(a/B ) does not fork over A. So tp(a/acleq (B )) does not fork over A. So p does not fork over A and hence is denable almost over A. Thus c acleq (A). (We are using transitivity, as well as the denition of nonforking here.) (iii) This will be something special about t.t. theories. Let p(x) S (A) be stationary, and c = Cb(p). So we know that p0 (x) = p|c is stationary (by (i)). By 2.29 (iv), p0 has Morley degree 1. Let (x, c0 ) be a formula over c which is in p0 and has Morley rank = RM (p0 ) and Morley degree 1. In fact there is no harm in allowing c0 to be a code for (x, c0 ). Let p(x) be the global nonforking extension of p0 . So RM (p) = and (x, c0 ) p. Let which xes c0 . Then f (p) has Morley rank f be an automorphism of M and contains (x, c0 ). Hence f (p) = p. Hence (by denition), f (c) = c. So . So c dcleq (c0 ) as all elements of c are xed by all c0 -automorphisms of M required. 23

(iv) Let p(x) S (A) be a type of Morley rank which contains (x). So by 2.29(iv) p is stationary. By the proof of (iii) above there is a formula (x) over c of Morley rank and degree 1 such that if c0 is a code for (x), then c = dcleq (c0 ). Note that as (x) (x) p(x), (RM, dM )((x) (x)) = (, 1) (hence, in passing, the symmetric dierence of and has Morley rank < ). Let p be the global nonforking extension of p(x). Note that for , f (p) = p i RM ( (x) f ( )(x)) = . But also any automorphism f of M f (p) = p if f (c0 ) = c0 (as Cb(p) = dcleq (c0 )). So putting it together we see , RM ( (x) f ( (x)) = , i f ( (x)) is that for any automorphism f of M equivalent to (x). Exercise 2.39 (i) Suppose that (bi : i < ) is an A-indiscernible sequence (of nite tuples of the same length). Then for all i, j < , stp(bi /A) = stp(bj /A). (ii) (T stable.) Suppose that (bi : i < ) is an A-indiscernible sequence of tules (of some xed nite length). Then {bi : i < } is totally indiscernible over A, namely for any n and n-element subsets {i1 , .., in }, {j1 , .., jn } of , tp(bi1 , .., bin /A) = tp(bj1 , .., bjn /A). Denition 2.40 (T t.t.) Let p(x) S (A) be stationary. By a Morley sequence (of length ) of p(x) we mean a sequence (bi : i < ) such that each bi realizes the (unique) nnonforking extension of p(x) over A {b0 , .., bi1 } (or in previous notation bi realises p|(A {b0 , .., bi1 }). Exercise 2.41 (T t.t.) Let (bi : i < ) be a Morley sequence of the stationary type p(x) S (A). Then (bi : i < } is A-indiscernible. Moreover if (ci : i < ) is another Morley sequence of p(x), then tp((bi )i< /A) = tp((ci )i< /A). Proposition 2.42 (T t.t) Let p(x) S (A) be stationary. Let (bi : i < ) be a Morley sequence of p(x). Then Cb(p(x)) dcleq (bi : i < ) (in fact Cb(p(x)) dcleq (b0 , .., bn ) for some n < ). Proof. Let p be the global nonforking extension of p(x). By inspecting the proof of Lemma 2.6, we see that for any (x, y ) L, the -denition of p(x) is over (ai : i < ) for some Morley sequence (ai )i of p(x). But this -denition is also over A. Hence by the second part of Exercise 2.41, it is over any Morley sequence (bi : i < ) of p(x). This shows that c = Cb(p) is in the denable closure of any Morley sequence of p(x). As (by 2.38 (iii)) c is 24

contained in dcl(c0 ) for a single imaginary c0 c, we obtain the parenthetical remark too. Denition 2.43 (T t.t.) Let {bi : i I } be a set of tuples, and A a set of parameters. We will say that {bi : i I } is independent over A or Aindependent, if for each i I , bi is independent from A {bj : j I, j = i} over A. Exercise 2.44 (T t.t) Suppose that (bi : i < ) is a sequence such that bi is independent from (b0 , .., bi1 ) over A for each i < . Show that {bi : i < } is A-independent. (Forking calculus.) So note in particular that any Morley sequence of a stationary type p(x) S (A) will be A-independent. Lemma 2.45 (T t.t). Suppose that A is a set of parameters, {bi : i I } is an A-independent set of tuples, c is a nite tuple, and for each i I , tp(c/Abi ) forks over A. Then I is nite. Proof. Suppose not. Without loss I = . Claim. For each i, tp(bi+1 /A {b0 , .., bi } {c}) forks over A {b0 , .., bi }. If not then by our assumptions, together with transitivity, tp(bi+1 /A{b0 , .., bi } {c}) does not fork over A, hence by transitivity again tp(bi+1 /A {c}) does not fork over A, so by symmetry, tp(c/A {bi }) does not fork over A, contradicting the hypotheses. By the claim and symmetry, tp(c/A {b0 , .., bi+1 }) forks over A {b0 , .., bi }) for all i < . By Proposition 2.26, we have RM (tp(c/A)) > RM (tp(c/A {b0 })) > ... > RM (tp(c/A {b0 , .., bi })) > RM (tp(c/A {b0 , .., bi , bi+1 })) > ... which gives a strictly descending innite sequence of ordinals, a contradiction. Denition 2.46 (i) Let (x, b) be a formula with parameter b. (x, b) is said to divide over A if there is an innite A-indiscernible sequence (bi : i < ) of realizations of tp(b/A) such that {(x, bi ) : i < } is inconsistent (namely ). not realized in M (ii) A complete type p(x) S (B ) is said to divide over A B , if some formula (x) p(x) divides over A. Example 2.47 (a) Suppose (x, b) is consistent and almost over A. Then (x, b) does not divide ove A. (b) Suppose that b / acl(A). Then the formula x = b divides over A. 25

Proof. (a) Let (bi : i < ) be A-indiscernible with tp(b/A) = tp(bi /A) for some (all) i. The formulas (x, bi ) are images of (x, b) under automorphisms which x A pointwise. So, as (x, b) is almost over A there must be i < j such that |= x((x, bi ) (x, bj )). So by indiscernibility, all (x, bi ) are equivalent. So clearly {(x, bi ) : i < } is consistent (as (x, b) is consistent). (b) As b / acl(A) there are realizations bi of tp(b/A) for i < such that i = j implies bi = bj . By Proposition 5.11 of [1], we may assume that (bi : i < ) is A-indiscernible. But then x = bi x = bj is inconsistent for i = j . This shows that x = b divides over A. Remark 2.48 Shelahs original denition of forking was as follows: (i) The formula (x, b) forks over A if there are formulas 1 (x, b1 ), ..., n (x, bn ) such that |= x((x, b) i i (x, bi )) and each i (x, bi ) divides over A. (ii) p(x) S (B ) forks over A B if there is a formula (x, b) p(x) which forks over A. The next proposition shows that our denition of forking agrees with Shelahs original one (for t.t theories). Lemma 2.49 (T t.t) Let p(x) S (B ) and A B . Then (i) p(x) forks over A (in the sense of Denition 2.24) if and only if p(x) divides over A. (ii) p(x) forks over A (in the sense of Denition 2.24) if and only if p(x) forks over A in Shelahs sense (in Remark 2.48) Proof. (i) Suppose that p(x) forks over A. By 2.29 (vi) there is nite b B such that p(x)|Ab forks over A. Write p(x)|Ab = q (x, b). Let r(y ) = stp(b/A). Let (bi : i < ) be a Morley sequence of r(y ). Claim. {q (x, bi ) : i < } is inconsistent. Proof. If not, there is c realizing it. By automorphism tp(c/Abi ) forks over A for all i. Now (bi : i < ) is A-independent (why?), so Lemma 2.45 gives a contradiction. By the claim and compactness there is a formula (x, b) q (x, b) such that {(x, bi ) : i < } is inconsistent. By 2.41, (bi : i < ) is A-indiscernible. So (x, b) divides over A. (Note that phi(x, b) may contain hidden parameters from A, say a. Write (x, b) as (x, b, a). So (x, b, a) p(x), (bi , a) : i < ) is an A-indiscernible sequence of realizations of tp(ba/A), and { (x, bi , a) : i < } is inconsistent, hence (x, b, a) divides over A.) 26

Conversely, suppose that p(x) does not fork over A. Let (x, b) p(x), and let (bi : i < ) be an A-indiscernible sequence of realizations of tp(b/A). We may assume that b = b0 . Let a realize p(x), let p (x) be tp(a/acleq (B )), and let p(x) be the unique global nonforking extension of p . Then p does not fork over A (by transitivity) hence is denable over acleq (A). In particular , (x, b ) p i there is a formula (y ) over acleq (A) such that for any b M eq eq |= (b ). By Exercise 2.39 (i), tp(bi /acl (A)) = tp(b/acl (A)) for all i < . As (x, b) p(x), we have |= (b), and thus |= (bi ) for all i < , and so (x, bi ) p(x) for all i. So {(x, bi ) : i < } is consistent. So (x, b) does not divide over A. (ii) Note that if (x) divides over A then (x) forks over A in the sense of Shelah, so LHS RHS follows from (i). Conversely, suppose p(x) S (B ) and p(x) forks over A B in the sense of Shelah. So there are nitely many formulas i (x) for i = 1, .., n say, such that p(x) |= i i (x), and each i (x) divides over A. Let C B contain the parameters from the i . So clearly any extension of p(x) to a complete type over C contains one of the i hence divides over A. On the other hand p(x) has, by 2.29, a nonforking extension (in sense of 2.24) extension q (x) S (X ). By what we just said, q (x) divides over A. By part (i), q (x) forks over A in the sense of 2.24. By 2.29 (transitivity), p(x) forks over A (in sense of 2.24). Here is another characterization of forking in t.t. theories. The proof makes use of 2.40 although it could be given without it. Lemma 2.50 (T t.t) Let p(x) S (B ) and A B . The following are equivalent: (i) p(x) does not fork over A, (ii) for any formula (x) p(x) and model M A, there is c M such that |= (c). Proof. Let us remark rst that the model M in (ii) need not contain B . (i) implies (ii). Assume p(x) does not fork over A. Let (x, b) p(x) (where we exhibit the parameters) and let M be a model containing A. Let p (x) S (M B ) be a nonforking extension of p(x). Then by transitivity, p (x) does not fork over M . Let a realizes p (x). By transitivity and symmetry, tp(b/M, a) does not fork over M . We then know that tp(b/M, a) is precisely d(M a) where d is a dening schema for tp(b/M ). Let (x) = d((x, y )). 27

So |= (a), hence (x) is consistent, hence realized in M by a say (as (x) is over M ). Thus |= (a , b). Conversely, assume (ii). We will show that p(x) does not divide over A. Let (x, b) p(x). Let (bi : i < ) be an A-indiscernible sequence of realizations of tp(b/A), where we may assume that b0 = b. Let M A be a model such that (bi : i < ) is also M -indiscernible. (We leave it as an exercise to be added to assignment 3, that such M can be found.) By assumption there is c M such that |= (c, b0 ). But then |= (c, bi ) for all i, hence {(x, bi ) : i < } is consistent. Finally let us return to the question of the existence of saturated models of t.t theories. We proved in Lemma 5.20 of [1] that if T is countable and -stable, then for any innite regular cardinal, T has a saturated model of cardinality . The proof goes over to the t.t. case: T has a saturated model of cardinality for every regular |T |. We can now deal with the general case. First: Remark 2.51 Suppose T is t.t. Then for every |T |, T is -stable, that is, for every model M of T of cardinality , there are only -many complete n-types over M for all M . Proof. A complete type p(x) over a model M is determined by the choice of a formula (x) p such that RM (p(x)) = RM ((x)) and dM ((x)) = 1. Proposition 2.52 Suppose T is t.t. Then for every |T |, T has a saturated model of cardinality . Proof. By Remark 2.51, we can build a continuous elementary chain (M : < ) of models of T of cardinality , such that all complete (nitary) types over M are realized in M+1 . Let M be the union of the chain. So M is of cardinality . We claim that M is -saturated. Let A M have cardinality < , and p(x) S (A). We must show that p is realized in M . By adding a nite set of parameters to A we may assume that p(x) is stationary (why?). Let A0 be a nite subset of A such that p(x) does not fork over A0 and p|A0 is stationary. (For example, let (x) p(x) have least Morley rank, and Morley degree 1 and let A0 be the parameters from .) Then A0 is contained in M for some < . Let p0 = p|A0 . By choice of the M we can nd c +1 M +1 for < , such that c +1 realizes the unique nonforking 28

extension of p0 (x) over M . Then (c ) is a Morley sequence in p0 (x) (of length ) in particular, is A0 -independent. By 2.45, for each nite tuple b from A all but nitely many c s are independent from b over A0 . By the nite character of forking, and the fact that A has cardinality < there is some c which is independent from M over A0 . But then tp(c /A) is the unique nonforking extension of p0 over A which is p(x). So p(x) is realized in M . In the next section we will discuss structures/denable sets of nite Morley rank, especially of rank 1, and there will be many examples. So we conclude this section with some innite Morley rank examples. Example 2.53 Let Ei for i < be binary relation symbols. Let Ln = {Ei : i n}. Let Tn be the Ln -theory with axioms: - for i n, Ei is an equivalence relation with innitely many classes, - E0 is equality, - if i < n then every Ei+1 -class is a union of innitely many Ei -classes. Let T = n Tn in language L = n Ln . Then in Tn , (RM, dM )(x = x) = (n + 1, 1), and in T , (RM, dM )(x = x) = (, 1). Explanation The proof of this is left to you. First show that each Tn is complete with quantier-elimination. Show that a formula Ei (x, a) has Morley rank i and degree 1. How can you describe forking for 1-types in these theories? Before the next example, recall that we proved in 3.6 of [1] that ACFp is complete with QE in the language of rings (for p a prime or zero). It follows from QE that any denable subset of an algebraically closed eld is nite or conite, hence x = x has Morley rank 1 and degree 1. In particular ACFp is -stable (why??). Example 2.54 DCF0 was introduced in Example 3.17 of [1], where we said (without proof ) that DCF0 is complete with quantier-elimination. We use this to sketch a proof that DCF0 is -stable, hence t.t.. Let (K, +, , , 0, 1, ) be a countable model of DCF0 and let K be a saturated elementary extension. We consider complete 1-types over K realized in 29

K and show there are only countably many of them. Let L be the language of dierential rings and Lr the language of rings. (K, +, , , 0, 1) will be an algebraically closed eld of characteristic 0. Fix p(x) S1 (K ), and let a K realize it. We have two cases: Case 1. (a, (a), 2 (a), ....) is algebraically independent over K . (Namely for any polynomial P (x0 , ...., xn ) K [x0 , .., xn ], P (a, (a), .., n (a)) = 0. By QE, p(x) is uniquely determined. That is, there is at most one p(x) S1 (K ) such that Case 1 holds. Case 2. Otherwise. Let n be least such that (a, (a), .., n (a)) is algebraically dependent over K . So n depends only on p(x) and we call n the order of p. Claim. p(x) is determined by n = ord(p) together with tpLr (a, (a), .., n (a)/K ). Let P (x) be the minimal polynomial over n (a) over K (a, (a), .., n1 (a)). Then applying to P ( n (a)) = 0, and using minimality of P (X ), one nds a K -rational function s(x0 , .., xn ) such that n+1 (a) = s(a, (a), .., n (a)). Continuing, one nds K -rational functions si (x0 , .., xn ) such that n+i (a) = si (a, (a), .., n (a)). The conclusion is that if b K , ord(tp(b/K )) = n and tpLr (b, (b), .., n (b)/K ) = tpLr (a, (a), .., n (b)/K ) then for all i 1, n+i (b) = si (b, ..., n (b)). It follows that for all m, tpLr (b, (b), .., m (b)/K ) = tpLr (a, (a), .., m (a)/K ), and so by QE, tp(b/K ) = tp(b/K ) = p(x), proving the claim. By the claim together with -stability of ACF0 , there are only countably many 1-types over K of nite order. Together with Case 1 giving a unique 1-type we see that there are only countably many complete 1-types over K . This proves -stability of DCF0 . Let us point out a few additional things about 1-types in DCF0 without proof: - RM (x = x) = , - k1 k2 K are dierential subelds of K , and p(x) is a complete 1-type over k2 , then p(x) does not fork over k1 i ord(p(x)) = ord(p(x)|k1 ).

Strongly minimal sets and geometry

Let T be a (complete) theory in a language L, and ( x) an L-formula. We say that ( x) is strongly minimal if RM () = 1 and dM () = 1. If X is 30

an A-denable set in some structure M we say that X is strongly minimal, if the formula dening X is a strongly minimal formula in the theory T h(M, a)aA . be an -saturated structure, and X M n a set which Remark 3.1 Let M , where A is a nite set of parameters from M . Then X is A-denable in M is strongly minimal if X is innite, and for every denable (with parameters) n , X Y is nite or X \ Y is nite. subset Y of M By a strongly minimal theory we mean a (complete) 1-sorted theory T in which the formula x = x is strongly minimal. Strongly minimal formulas and sets are important for t.t theories of nite Morley rank. Any structure of nite Morley rank can be more or less built out of strongly minimal sets. For groups of nite Morley rank this can be made more precise: if G is a group of nite Morley rank, then there are normal denable subgroups {1} < N1 < .. < Nk = G such that each quotient Ni /Ni1 is almost strongly minimal. Studying strongly minimal sets amounts to the same thing as studying strongly minimal theories. Let us rst elaborate on this, as it gives us an opportunity to discuss induced structure and stable embeddedness. First: Remark 3.2 Let T be t.t. Then any complete type over any set A is denable: for each (x, y ) L there is (y ) LA such that for b A, (x, b) p(x) i |= (b). Proof. Let (x) p(x) have least (RM, dM ), say (, d). So for any (x, y ) L and b A, (x, b) p(x) i (RM, dM )((x) (x, b)) = (, d). Now use Corollary 2.15. Denition 3.3 Let M be an L-structure, possibly many-sorted. Let S be a sort of L, and X a denable subset of S (M ), dened with parameters A from M . Then by the induced structure over A on X we mean the following: The 1-sorted language L consists of an n-place relation symbol RY for each A-denable subset Y of S (M )n . Make X into an L -structure with universe X by interpreting each RY as Y X n . Lemma 3.4 Assume T to be t.t. Let M be a model of T (maybe saturated). Let X be an A-denable set in M . Let X be the L -structure on X given by 31

the denition above. Then (i) T h(X ) has quantier-elimination in L . (ii) The subsets of X n denable (with parameters) in X are the same as those denable (with parameters) in M . Proof. (i) is obvious. (ii). By (i) any subset of X n denable with parameters in X is denable (with the same parameters) in M . Conversely, let Y X say be denable by the formula (x, b) in M where b may be outside X . By Remark 3.2, tp(b/X ) is denable (over X ). So there is a formula (x, c) with parameters c from X , such that for a X , M |= (a, b) i M |= (a, c). Let Z be the set in M dened by (x, z ). Then Y is dened in X be the formula RZ (x, c). Remark 3.5 Note that, in Lemma 3.4, if M is 1-sorted and X is M itself, then the structure X is essentially just the Morleyization of M . Let us consider some examples. The rst has been mentioned before: Example 3.6 The theory of algebraically closed elds of some xed characteristic is strongly minimal. Proof. ACFp has quantier-elimination in the language of rings. So if K is an algebraically closed eld, every denable (with parameters) subset of K is a Boolean combination of sets dened by things of the form P (x) = 0 where P K [x] is a polynomial over K in single indeterminate x. The set of solutions of P (x) = 0 is either nite or everything. So any denable subset of K is nite or conite. Example 3.7 The formula (x) = 0 is strongly minimal in DCF0 . Moreover the induced structure on the corresponding strongly minimal set in a model of DCF0 is just that of an algebraically closed eld. Proof. Because of 3.4(ii) and 3.6 it is enough to show the second part. Let K be a dierentially closed eld. The solution set of (x) = 0 in K is usually called the constants of K and sometimes dentoted CK . I will be using the fact that CK is algebraically closed. (Proof: A derivation on a eld F extends uniquely to a derivation on the algebraic closure of F . Thus every existentially closed dierential eld K is algebraically closed. By the uniqueness assertion above the algebraic closure in K of CK must also consist of 32

constants.) As CK is dened without parameters, we take A as in 3.3 to be . n Let X now be a -denable subset of K n , and consider X CK , and we want to show that this latter set is -denable in (CK , +, , 0, 1). By quantierelimination in DCF0 we may assume that X is given by P ( x, x , ...) = 0, n where P is a polynomial over Z. As is 0 on CK , X CK is dened by P ( x, 0, 0, ..) = 0. n (The point we are making here is that the subsets of CK which are denable in the structure (K, +, , , 0, 1, ) are the same as those which are -denable in the structure (CK , +, , , 0, 1). The theory of the latter structure is strongly minimal, by 3.6. By 3.4 (ii), and t.t.-ness of DCF0 , CK is a strongly minimal set in K .) Example 3.8 Let F be a division ring. Let V be an innite-dimensional vector space over F . Consider V as a structure in the language containing +, , 0, as well as unary functions f for f F (representing scalar multiplication by elements of F ). Then T h(V ) is strongly minimal. Proof. A back-and-forth argument shows that T h(V ) has quantier-elimination in the given language. So, up to Boolean combination, a formula (x, a1 , .., an ) is of the form x + i=1,..,n i ai = x + i=1,..,n i ai which is equivalent to x = b for some b V . Good. (Note that every denable subset of V n is, up to Boolean combination, of the i=1,..,n i xi = b.) of T and a strongly Let us now x a t.t. theory T , a saturated model M minimal set D -denable in M . The reader can assume D to be precisely if he or she so wishes. We will study denability and independence in D. M x will denote a variable of the sort of which D is a subset. Here is a special case of Lemma 2.14, which has an easier proof. Remark 3.9 Let (x, y ) be an L-formula. Then there is an L-formula (y ) , |= (b) i (x, b) D(x) is nite. such that for all b M such that both Proof. If not, then for arbitrarily large n there is b M (x, b) D(x) and (x, b) D(x) have at least n solutions. By compactness there is b such that both (x, b) D(x) and (x, b) D(x) have innitely many solutions, contradicting strong minimality of D.

33

Insofar as denable subsets of D itself are concerned, the innite ones have Morley rank 1 and the nite ones have Morley rank 0. Let us take an informal look at denable subsets of D D. Let X D D be denable. Let be the projection on the rst coordinate. For a D, let Xa be the bre {b : (a, b) X }. Let Y = {a D : Xa = }, Y1 = {a Y : Xa is nite} and Y2 = {a Y : Xa is innite}. Y is clearly denable, and Y1 , Y2 are denable by Remark 3.9. Let Xi = 1 (Yi ) X for i = 1, 2. So X is the disjoint union of X1 and X2 . If Y1 is nite, then so is X1 . If Y1 is innite (so conite), then X1 is intuitively something 1-dimensional. If Y2 is nite and nonemmpty, then again X2 is something 1-dimensional. If Y2 is innite (so conite) then intuitively X2 is 2-dimensional. As we shall see these dimensions correspond exactly to Morley rank. We will need the following which is related to Exercise 2.2 (ii). Exercise 3.10 Suppose that a, b are nite tuples and b acl(A, a). Then RM (tp(a/A)) RM (tp(b/A)) and RM (tp(a/A)) = RM (tp(a, b/A)). Let p0 be the unique complete type over which contains D(x) and has Morley rank 1 (and degree 1). Note that p0 (x) is axiomatized by {D(x)} x / acl(). Also p0 is stationary. . The following are equivalent: Remark 3.11 Let a D and A M (i) a / acl(A), (ii) a realizes p0 |A. Proof. The nonforking extensions of p0 are by 2.26 the extensions of p0 of Morley rank 1. But the types of Morley rank 0 are precisely the algebraic types. Corollary 3.12 Let a, b D and A a set of parameters. Suppose b acl(A, a) \ acl(A). Then a acl(A, b) \ acl(A). Proof. This follows from forking symmetry for t.t theories, bearing in mind 3.11. Of course Corollary 3.12 can be proved directly and easily, without recourse to forking symmetry in t.t. theories. 34

Let us make a quick denition, which makes sense in any structure: Let , and A a set of {bi : i I } be a set of nite tuples in a structure M parameters. We call {bi : i I } algebraically independent over A if for each i, b i / acl(A {bj : j I, j = i}). Back in our situation: Corollary 3.13 Let {bi : i < } D. Assume that bi / acl(A {bj : j < i}) for all i < . Then (i) {bi : i < } is an A-independent set of realizations of p0 /A, and (ii) {bi : i < } is algebraically independent over A Proof. (i) is a direct application of 3.11 and 2.44. (ii) follows from (i) and 3.11. The next result will show that Morley rank equals dimension inside strongly minimal sets. Proposition 3.14 Let a = (a1 , .., am ) be an m-tuple of elements of D. Assume that (a1 , .., an ) is algebraically independent over A, and that a acl(A {a1 , .., an }). Then RM (tp( a/A)) = n. Proof. We prove this by induction on n. For n = 1, RM (tp(a1 /A)) = 1 by 3.11, and then RM (tp( a)) = 1 by 3.10. Now assume true for n < n, and we prove it for n. By 3.10 it is enough to prove that RM (tp(a1 , .., an /A)) = n. Now by induction hypothesis RM (tp(a1 , .., an /A {a1 })) = n 1 and by 2.47(b) and symmetry, tp(a1 , .., an /A {a1 }) forks over A. So by 2.26, RM (tp(a1 , .., an /A)) n. Suppose for a contradiction that RM (tp(a1 , .., an /A)) > n. So RM (Dn ) > n. It follows that there are at least two complete types q1 (x1 , .., xn ) and q2 (x1 , .., xn ) over some set B , which contain x Dn , and realize q2 . By the induction both have Morley rank n. Let c realise q1 and d hypothesis, each of the sets {c1 , .., cn }, and {d1 , .., dn } is algebraically inde is an B -independent set of realizations pendent over A. By 3.11, each of c , d ), of p0 |B and is thus a Morley sequence in p0 |B . By 2.41, tp( c/B ) = tp(d/B a contradiction. This proves that RM (tp(a1 , .., an /A)) = n, completing the proof of the proposition. and B D. By a basis of B over A we mean Denition 3.15 Let A M a maximal subset B0 of B such that B0 is algebraically independent over A. 35

Remark 3.16 Let B0 be a basis of B over A. Then B acl(A B0 ). Proof. If not let b B such that b / acl(A B0 ). By Corollary 3.13, B0 {b} is algebraically independent over A, contradicting maximality of B0 . Lemma 3.17 (With the notation of Denition 3.15.) Let B0 and B1 be two bases of B over A. Then B0 and B1 have the same cardinality. Proof. Suppose rst that one of the bases is innite. Without loss of generality |B0 | < |B1 | = and is innite. For each element b B1 there is a nite tuple cb from B0 such that b acl(A cb ). It follows that there is some nite tuple c from B0 and and innite subset B0 of B0 such that b acl(A c) for all b B0 . So tp(b/A c) forks over A for all b B0 and (by 3.13), B0 is A-independent. This contradicts Lemma 2.45. n Now assume that both B0 and B1 are nite. Let (b1 0 , .., b0 ) be an enumer1 m ation of B0 and (b1 , .., b1 ) an enumeration of B1 . Let c be the concatenation of these two tuples. Then By 3.14 and 3.16, n = RM (tp(c/A)) = m. Remark 3.18 (With notation as above.) By dim(B/A) we mean the cardinality of a basis for B over A. If B is nite and b is an enumeration of B , we write dim(b/A). So note, by 3.14 and 3.17, if b is a nite tuple from D then dim(b/A) = RM (tp(b/A)). Lemma 3.19 Let b, c be nite tuples from D. Then dim(bc/A) = dim(b/A c)+dim(c/A). Hence by Remark 3.18 and 3.14, RM (tp(bc/A)) = RM (tp(b/cA))+ RM (tp(c/A)). Proof. If b0 is a basis for b over A c and c0 a basis for c over A, then b0 c0 is a basis for bc over A. Exercise 3.20 Morley rank is denable in strongly minimal sets. That is, for any formula (x1 , .., xn , y ) (where y is some tuple of variables and implies that each xi is in D), and for any m, there is an L-formula (y ), (of length that of y ), RM ((x1 , .., xn , b)) = m such that for all tuples b in M i |= (b).

36

 eq : for some nite tuple d Denition 3.21 (i) By Deq we mean {c M from D, c dcl(d)}. eq , DA eq : for some nite tuple d from D, c (ii) For A M = {c M dcl(A d)}. (iii) Daeq and Daeq are dened as in (i) and (ii) but with acl() replacing dcl(). eq , dened over A say. We say that X is (iv) Let X be a denable set in M eq almost strongly mininimal (with respect to D) if for some B A, X DB . Remark 3.22 Let c be a nite tuple from D and A any set of parameters . Then Cb(stp(c/A)) Deq . from M Proof. By Proposition 2.42. Lemma 3.23 Let c Daeq . Then there is some set A of parameters and some nite tuple d from D such that c is independent from A over , and c is interalgebraic with d over A. Hence RM (tp(c/)) = RM (tp(c/A)) = RM (tp(d/A)). = (d1 , .., dn ) be some tuple from D such that c acl(d1 , .., dn ). Proof. Let d = (d1 , .., dr ) be a maximal subtuple of d which Relabelling, if necessary, let d ) is independent from c over . It follows that for i = 1, .., n, di acl(c, d , and d to be d . (why?). Take A to be d
aeq ). Then Corollary 3.24 . Let A be any set, and let b, c Daeq (or even DA RM (tp(bc/A)) = RM (tp(b/cA)) + RM (tp(c/A)).

Proof. Let us add names for elements of A to the language. We have to show that RM (tp(bc/)) = RM (tp(b/c)) + RM (tp(c/)). By the lemma above, let B be independent from b such that b is interalgebraic with some nite tuple from D over B . Likewise, let C be independent from c such that c is interalgebraic over C with a nite tuple from D. We may choose B, C such that (b, c) is independent from B C over . (First let B realise a nonforking extension of tp(B/b) over (b, c). Then let C realize a nonforking extension of tp(C/c) over B {b, c}. Replace B, C by B , C .) Let E = B C . So Let d1 be a nite tuple from D interalgebraic with b over E , and d2 a nite tuple from D interalgebraic with c over E . Then by Exercise 3.10, RM (tp(b, c/E )) = RM (tp(d1 , d2 /E )), RM (tp(c/E )) = RM (tp(d2 /E )) 37

and RM (tp(b/cE )) = RM (tp(d1 /d2 E )). On the other hand, as (b, c) is independent from E over (and thus also b is independent from cE over c), RM (tp(b, c/E )) = RM (tp(b/c)), RM (tp(b/cE )) = RM (tp(b/c)) and RM (tp(c/E )) = RM (tp(c/)). Together with 3.19, this yields the required result. The material above, especially 3.17, can be developed at a greater level of generality, that of pregeometries and geometries. I will say a little about this now. Let S be a set and cl a closure operator on S , namely a map from the power set P (S ) of S to itself. The pair (S, cl()) is said to be a (combinatorial) pregeometry if the following hold (where X, Y denote subsets of S and a, b elements of S ): (i) X cl(X ), (ii) X Y implies cl(X ) cl(Y ). (iii) cl(cl(X )) = cl(X ), (iv) If a cl(X {b}) \ cl(X ) then b cl(X {a}). (v) If a cl(X ) then a cl(X ) for some nite subset X of X . In a pregeometry we obtain a notion of independence: X S is independent if for each x X , x / cl(X \ {x}). Any subset X of S contains a maximal independent subset X0 , which we call a basis of X . All bases of X have the same cardinality, which we call dim(X ). If A is a subset of S , then the localization of S at A is the pregeometry which has the same underlying set S , but a new closure operation clA () where clA (X ) = cl(A X ). We write dim(X/A) for dim(X ) in the localization of S at A. Note that (vi) dim(X Y ) = dim(X/Y ) + dim(Y ). There are various rather important properties which a pregeometry (S, cl) may or may not have: (a) Triviality: cl(X ) = xX cl({x}), for all X . (b) Modularity: dim(X ) + dim(Y ) dim(X Y ) = dim(X Y ) for all closed X and Y . (c) Local modularity: The localization of (S, cl) at some singleton of S is modular. (d) Local niteness: for nite X , cl(X ) is nite. (e) Homogeneity: for any closed X S and a, b S \ X , there is an 38

automorphism of S (that is a permutation of S preserving cl()) which xes X pointwise and takes a to b. Exercise 3.25 (i) The pregeometry (S, cl) is modular i and only if property (b) above holds for all closed X, Y S such that dim(X ) = 2. (ii) (S, cl) is modular i for all closed X, Y S , dim(X/Y ) = dim(X/X Y ). Hint. For the second part use property (vi) above. Let us now return to the strongly minimal subset -denable subset D of the of the t.t theory T . saturated model M Lemma 3.26 Let cl() be algebraic closure restricted to D. Then (D, cl) is a homogeneous pregeometry. Proof. (i), (ii), (iii) and (v) in the denition of a pregeometry are clear. (iv) (although it is is by 3.12. For homogeneity, we will assume that D = M true in general). Let X be an algebraically closed subset of D (even of big cardinality), and let a, b D \ X . Extend each of a, b to bases A, B of D over X . Then A and B have the same cardinality, by 3.17. Enumerate A as (ai : i < ) and B as (bi : i < ) where a = a0 and b = b0 . Then these two sequences have the same type over X . The elementary map which xes X pointwise and takes (ai : i < ) to (bi : i < ) extends (by taking algebraic closures) to an automorphism of the structure D. Remark 3.27 (D, acl) is modular if and only for for any tuples a, b D, a is independent from b over acl(a) acl(b) D (in the sense of nonforking). Proof. Suppose (D, acl) is modular. It is enough to prove the RHS for a, b nite tuples from D. By 3.25 (ii), dim(a/b) = dim(a/acl(a) acl(b) D). So by 3.18, RM (tp(a/b)) = RM (tp(a/acl(b) D)) = RM (tp(a/acl(a) acl(b) D)), which is enough. Conversely, in the same way the right hand side implies that dim(a/b) = dim(a/acl(a) acl(b) D)) for all tuples from D, and we can use 3.25 again. = D = an innite set in the empty language. Example 3.28 (i) Let M Then (D, acl()) is a trivial pregeometry. 39

 = D = an (innite-dimensional) division vector space over a (ii) Let M division ring (as in Example 3.8). Then (D, acl) is modular. = D = an algebraically closed eld in the language of rings. Then (iii) Let M (D, acl) is not locally modular. Proof. Exercise. a saturated model and X a -invariant Denition 3.29 Let T be t.t, M eq subset of M . We say that X is 1-based if for any nite tuple a of elements eq such that tp(a/A) is stationary, the canonical base of X and any A M of tp(a/A) is contained in acleq (a). eq , Exercise 3.30 X is 1-based i for any tuple a from X , and any A M a is independent from A over acleq (a) acleq (A). Hint. Use 2.38 (ii). From Remark 3.27 and Exercise 3.30, we see some kind of formal similarity between the notions of modularity and 1-basedness. Lemma 3.31 (i) Suppose X is 1-based. Then so is X aeq . (ii) 1-basedness of X is invariant under naming parameters. Proof. We will use the fact (exercise) that if a, b, are tuples and a is independent from b over C then acl(a) acl(b) acl(C ). (i) Let d be a nite tuple from X and let c acl(d) and consider a stationary type tp(c/A) where without loss, A is eq -algebraically closed. We may assume that d is independent from A over c. Hence by the fact mentioned above (*) acl(d) A acl(c). By 1-basedness of X , d is independent from A over acl(d) A. As c acl(d), c is independent from A over acl(d) A and thus over acl(c) A by (*). So Cb(tp(c/A)) acl(c) by 2.38. (ii) If X is 1-based, it clearly remains 1-based after naming parameters. Con , a)aA . Let c X and B some set versely, suppose X is 1-based in T h(M such that tp(c/B ) is stationary. We may assume that (c, B ) is independent . In particular, tp(c/BA) does not fork over B . So from A over in M c0 = Cb(tp(c/B )) = Cb(tp(c/BA)). Using our assumptions c0 acl(cA) B . But B is independent from cA over c, hence (by the fact above) c0 acl(c). 40

Lemma 3.32 Suppose D is modular. Then D is 1-based. Proof. Assume modularity of D. Let a be a nite tuple from D and A some , such that tp(a/A) is stationary. Let c = Cb(tp(a/A)). We know set from M (2.42) that c Deq , so there is some C D such that c dcl(C ). We may choose C so that C is independent from a over c and so therefore that a is independent from C over c. We may assume that C is algebraically closed (in D). Note that c = Cb(tp(a/C )) too. By 3.27, a is independent from C over acl(a) C . By 2.38, c acleq (acl(a)) = acleq (a). , Denition 3.33 D is said to be linear, if whenever a, b D, A M RM (tp(a, b/A)) = 1, and c = Cb(stp(a, b/A)), THEN RM (tp(c/)) 1. A remark: If a, b D and RM (tp(a, b/A)) = 1, then dim(a, b/A)) = 1 and we can consider (a, b) as a generic point of a plane curve X D D, which is dened over c. The rank of tp(c) measure the dimension of the family (Xc )c . Lemma 3.34 D is linear if and only if for any a, b D, and any set A, Cb(stp(a, b/A)) acl(a, b). Proof. Suppose rst that D is linear. Let c = Cb(stp(a, b/A)). If RM (tp(a, b/A)) = 2 then (a, b) is independent from A over , so c acl(). If RM (tp(a, b)) = 0, then c = (a, b). So we may assume that RM (tp(a, b/A)) = 1. If RM (tp(a, b/)) = 1 then again c acl(). So we may assume that RM (tp(a, b/)) = 2. So (a, b) forks with c over , whereby c / acl(), so by linearity, RM (tp(c)) = 1. Note that RM (tp(a, b/c)) = 1. By 3.24, RM (tp(a, b/c)) + RM (tp(c)) = RM (tp(c/a, b)) + RM (tp(a, b)). So 1 + 1 = RM (tp(c/a, b)) + 2, whereby c acl(a, b) as required. The converse follows by a similar computation. Theorem 3.35 The following are equivalent: , a)aA . (i) There is some set A such that D is modular in T h(M (ii) D is 1-based, (iii) D is linear, (iv) D is locally modular.

41

Proof. (i) implies (ii): By 3.32, D is 1-based after adding names for elements of A. By 3.31, D is 1-based. (ii) implies (iii): By 3.34. (iii) implies (iv). Let a0 D \ acl(). We will show that the localization of D at {a0 } is modular. This means precisely that D is modular after adding a constant for a0 to the language. By Exercise 3.25(i) it is enough to show that if a1 , a2 D are such that {a0 , a1 , a2 } is algebraically independent over , and A D is algebraically closed and is such that a0 A and dim(a1 , a2 /A) = 1, and ai / A for i = 1, 2 then acl(a0 , a1 , a2 ) A \ acl(a0 ) is nonempty. Let c = Cb(stp(a1 , a2 /A)). So c acleq (A). By linearity and 3.34, c acl(a1 , a2 ). Note that dim(a1 , a2 /c) = 1, and ai / acl(c) for i = 1, 2. So a2 acl(c, a1 ). As a0 / acl(a1 , a2 ), a0 / acl(c). Hence a0 realizes p0 |c. But a1 also realizes p0 |c. Hence tp(a0 , c) = tp(a1 , c). Let a2 be such that tp(a0 , a2 , c) = tp(a1 , a2 , c). Then a2 acl(c, a0 ). Thus a2 A \ acl(a0 ). Also a2 acl(a0 , a1 , a2 ). This completes the proof of (iii) implies (iv). (iv) implies (i) is immediate. Let us show that algebraically closed elds are NOT locally modular. Fix an algebraically closed eld (K, +, , , 0, 1) which we assume to be of uncountable cardinality say (hence -saturated). We know K to be strongly minimal. Let a, b K . Let C be the {(x, y ) K K : y = ax + b}. Then C is also strongly minimal (as it is in denable bijection with K ). Let c / acl(a, b). Let d = ac + b. Then tp(c, d/{a, b)}) has Morley rank 1 and degree 1. In fact (c, d) is a generic point of C . CLAIM. (a, b) is (namely is interdenable with) Cb(tp((c, d)/{a, b})). Proof of CLAIM. Let p(x, y ) be the global nonforking extension of tp((c, d)/{a, b}). We have to show (see Denition 2.34) that for any automorphism f of K , f (a, b) = (a, b) if and only if f (p) = p. Clearly left to right is true. Now suppose that f (a, b) = (a , b ) = (a, b). Write C as C(a,b) . Then C(a ,b ) C(a,b) is either empty, or is a singleton. As p has Morley rank 1, and (x, y ) C(a,b) p it follows that (x, y ) inC(a ,b ) p. Hence f (p) = p. So choosing a, b algebraically independent, we see that RM (tp(Cb(tp(c, d/{a, b})))) = 2, so K is not linear, hence not locally modular.

42

Let us briey discuss orthogonality and related notions. We are working in of the t.t. theory T . the saturated model M Denition 3.36 (i) Let p(x) S (A) and q (y ) S (B ) be stationary types. p is said to be nonorthogonal to q if there is C A B , and there are realizations a of p|C and b of q |C such that tp(a/C, b) forks over C . (ii) Let D1 be a strongly minimal set dened over A, and D2 a strongly minimal set dened over B . Let p(x) S (A) be the generic type of D1 (namely p(x) = {x D1 } x / acl(A), and likewise for q (y ). We say that D1 is nonorthogonal to D2 if p(x) is nonorthogonal to q (y ). Example 3.37 Let T be theory of disjoint innite unary predicates P and Q. Then P and Q are strongly minimal and orthogonal. Lemma 3.38 Let D1 , D2 be strongly minimal sets. Then D1 is nonorthogonal to D2 if and only if there is some denable R D1 D2 , such that the projections of R on D1 and D2 are innite, and for every x there are at most nitely many y such that R(x, y ) and dually. Proof. If D1 and D2 are nonorthogonal, then there is some C over which both are dened, and a D1 \ acl(C ) and b D2 \ acl(C ) such that a forks with b over C (why?). This implies that a acl(b, C ) and b acl(C, a). Let (x, y ) be a formula over C such that |= (a, b), and such that |= xk y(x, y ) and |= y l y(x, y ) for some k, l. Then (x, y ) denes the required relation R. Conversely, suppose such R exists. Let D1 , D2 , R be dened over C . By the assumptions on R there is (a, b) R such that a D1 \ acl(C ), and b D2 . Then b acl(C, a) and a acl(C, b). Hence tp(a/C, b) forks over C . Corollary 3.39 Nonorthogonality between strongly minimal sets is an equivalence relation. What became known as Zilbers conjecture was: (ZC1) Suppose D is strongly minimal and non locally modular. Then there such that is a strongly minimal algebraically closed eld D2 denable in M D1 is nonorthogonal to D2 . Zilbers conjecture has been very inuential. A counterexample was found by Hrushovski in the late 1980s, and the techniques used to construct it 43

gave rise to a whole subarea of model theory, Hrushovski constructions. Nevertheless the conjecture is true in many natural situations, and often has quite profound mathematical meaning and implications. An equivalent formulation of Zilbers conjecture is: (ZC2) Let X be a denable set of nite Morley rank. Suppose that X is not 1-based. Then there is a strongly minimal algebraically closed eld denable in X eq . Recent work has led to a somewhat stronger version of the Zilber conjecture which more or less states that any violation of 1-basedness lives in a denable eld in some sense. The notion of internality is important here. Denition 3.40 Let p(x) S (A) be stationary and X some B -denable set, or even some B -invariant set. We say that p(x) is (almost) internal to X if there is C A B and a realizing p(x)|C and some nite tuple c of elements of X such that a dcl(C, c) (a acl(C, c)). The strong version of Zilber conjecture then says: (ZC3) Suppose tp(a) has nite Morley rank, tp(a/A) is stationary and c = Cb(tp(a/A)). Then there is some denable set X which is a nite union of strongly minimal algebraically closed elds, such that stp(c/a) is almost internal to X . There is a weaker version of (ZC3) which does not mention elds and may be true. (NMC) Suppose tp(a) has nite Morley rank, tp(a/A) is stationary and c = Cb(tp(a/A)). Then there is some denable set X which is a nite union of non locally modular strongly minimal sets, such that stp(c/a) is almost internal to X . Let us nish by discussing internality a bit more. Lemma 3.41 Suppose p(x) S (A) is stationary and internal to the Ainvariant set X . Then there is a set B A, such that for every realization eq ) XB . a of p there is some tuple c from X such that a dcl(B, c). So p(M Proof. Let B A and a realize p|B and c X be such that a dcl(Bc). We may assume tp(a, c/B ) = q (x, y ) is stationary. Let (ai , ci )i< be a Morley 44

sequence in tp(a, c/B ). Let q (x, y ) be the nonforking extension of q (x, y ) over B {(ai , ci )}i , and let (a , c ) realize q . By 2.42, q does not fork over (ai , ci )i and the restriction of q to this set is stationary. It follows (why?) that a dcl(A (ai , ci )i c ). Thus there is n and nite d X such that a dcl(A (a1 , .., an ) d). Note that a is independent from (a1 , .., an ) over A (why?) Now let a be any realization of p(x). Let (a1 , .., an ) be such that tp(a , a1 , .., an /A) = tp(a , a1 , .., an /A) and (a1 , .., an ) is independent from (a1 , .., an ) over A a . In particular (a) there is d X such that a dcl(a1 , .., an , d , A), and (b) each ai is independent from (a1 , .., an ) over A, so has the same type over A{a1 , .., an } as a , whereby there is di X such that ai dcl(A, a1 , .., an , di ). By (a) and (b), a dcl(A, a1 , .., an , d , d1 , .., dn ). Note that it follows from 3.41 and compactness that under the same assumptions, there is an A-denable set Y in p(x), of maximal Morley rank, and a B -denable function from some B -denable subset of X n onto Y . Internality can give rise to denable groups if additional parameters are really needed to see it. This idea again originates with Zilber. Here is a rather striking case, which for now we state without proof. Lemma 3.42 Suppose that tp(a/A) is stationary and internal to the Ainvariant set X . Suppose also that a is independent from c over A for all c X (that is, a is independent from X over A). THEN there is an A ). denable group G and an A-denable transitive action of G on p(M

-stable groups

remains a saturated model of the t.t theory T and we will In this section M , sometimes specializing to denable groups of study denable groups in M nite Morley rank. eq without mentioning it. We will also work in M ) we mean a denable set G equipped with a By a denable group (in M denable group operation . If both the set G and the group operation are dened over A, we say that the group (G, ) is dened over A. Note then that both the identity of the group as well as group inversion are also A-denable.

45

We often just write G rather than (G, ) for the group. A special case is when is equipped with a -denable group operation. the universe of M In any case when we think of a denable group as a structure in its own right, it is not just the group structure but all the structure on G induced , in the sense of Denition 3.3. from M This is in accordance with the idea of an algebraic group, which is our basic example. An algebraic group (over a given algebraically closed eld K ) is an algebraic variety G over K together with a morphism G G G over K which is a group operation. Any algebraic group is a denable group in (K, +, ), and conversely it can be shown that any group denable in (K, +, .) is denably isomorphic to an algebraic group. There are two extremes of algebraic groups; abelian varieties (whose underlying variety is a Zariski closed subset of some projective space over K ) and linear algebraic groups, which are algebraic subgroups of some GL(n, K ), the group of invertible n n matrices over K . We may sometimes also consider type-denable groups, namely where the underlying set of the group is dened by a partial type. We will be using throughout the fact that the Morley rank and degree of a denable set are invariant under denable bijection. Lemma 4.1 Let (G, ) be a denable group, and H a denable subgroup of G. Then either RM (H ) < RM (G), or dM (H ) < dM (G). Proof. The left cosets a H of H in G partition G, and are are all in denable bijection with H so have the same Morley rank and degree as H . So if the index of H in G is innite, then RM (H ) < RM (G), and if the index is nite, n say, then RM (H ) = RM (G) and dM (G) = n(dM (H )). Corollary 4.2 A denable group G has the DCC on denable subgroups. (There is no innite descending chain of denable subgroups.) In particular, G has a smallest denable subgroup of nite index in G, which we call the connected component of G and write as G0 . G0 is normal in G and is Adenable if G is A-denable. Moreover RM (G0 ) = RM (G). Proof. The DCC is immediate from Lemma 4.1. It follows that G0 exists and is the intersection of all denable subgroups of nite index in G. By this property, G0 is invariant under A-automorphism of G hence A-denable. As any conjugate of a nite index subgroup also has nite index, G0 is normal. 46

We will tend to assume that our ambient denable group G is -denable in . M Denition 4.3 Assume that RM (G) = . (i) A denable subset X of G is said to be generic if RM (X ) = . (ii) Let p(x) S (A) be a complete type over a set A such that p(x) |= x G. p(x) is said to be a generic type of G (over A) if RM (p(x)) = . (usually a subset of G), and a G. We say (iii) Let A be any subset of M that a is a generic element of G over A if tp(a/A) is a generic type of G. Let us now x the -denable group G with Morley rank and Morley degree d. Note that there exist generic types of G over any set A. Note that also that if p(x) is a complete type over A of an element of G and q (x) is a nonforking extension of p then p is generic i q is generic. Also note that as inversion is an -denable bijection between G and itself, tp(a/A) is generic i tp(a1 /A) is generic. Finally note that a denable subset X of G is generic if and only if it is contained in some generic type of G. Let us begin with a characterization of generic types, with can be taken as the denition in the stable (or even simple) case. . The following are equivalent: Lemma 4.4 Let a G and A M (i) tp(a/A) is generic, (ii) if b G and a is independent from b over A then b a is independent from b over A, (iii) If b G, and a is independent from b over A then a b is independent from b over A. Proof. Assume (i). We will prove (ii) and (iii). Let b G be such that a is independent from b over A. So RM (tp(a/A, b)) = . So RM (tp(b a/A, b)) = . So (as = RM (G)), RM (tp(b a/A)) = , hence b a is independent from b over A. So we proved (ii). (iii) follows similarly. Now assume (ii). Choose b G generic over A {a}. Note that b is independent from a over A. So by (i) (iii), tp(b a/A) is generic. Note that a is independent from b over A. By (ii), b a is independent from b over A. Hence tp(b a/A, b) is generic, hence tp(a/A, b) is generic. As a is independent from b over A, tp(a/A) is generic. So we have (i). In a similar fashion (iii) implies (i).

47

By 2.33 (ii) there are d disjoint acl()-denable subsets of G, each of Morley rank , and Morley degree 1. So we obtain d distinct stationary generic types p1 (x), .., pd (x) of G, over acl(). Note that any stationary generic type of G is a nonforking extension of one of the pi . We aim towards showing that d is precisely the index of G0 in G. Let S = {p1 , .., pd }. Let us start by dening an action of G on S . Lemma 4.5 Let g G. Let a realize pi with a independent from g over . Then tp(g a/acl()) = pj for some j . Moreover j depends only on g and i (not on the choice of a). So we write g pi = pj . Proof. Note that RM (tp(a/g )) = , hence RM (tp(g a/g ) = , hence RM (tp(g a/acl()) = , so must be pj for some j {1, .., d}. As pi is stationary, it has a unique nonforking extension over acl() {g }, hence tp((g, a)/acl()) is uniquely determined by g and pi , and therefore also tp(g a/acl(). Exercise 4.6 The map from G S to S dened by (g, pi ) g pi is a group action. Lemma 4.7 (i) G acts transitively on S . (ii) for any i, j , {g G : g pi = pj } is denable (over acl()). Proof. Let pi , pj S . Let a, b realizes pi , pj respectively such that a is independent from b over . In particular b is independent from a over so independent from a1 over . By 4.4 and the genericity of tp(b), b a1 is independent from a over . Put g = b a1 , and we see that g.a = b, hence g pi = pj . (ii) Let i (x) be a formula over acl() of Morley rank and Morley degree 1 which implies x G, and is contained in pi (x). Let pi (x) be the global nonforking extension of pi . Now x i and j . Note that pj is denable (over acl()), hence {g G : i (g 1 .x) pj }, is denable, by a formula i,j (y ) over acl(). Claim. Let g G. Then |= i,j (g ) i g pi = pj . Proof of claim. Suppose rst that |= i,j (g ). So we can nd b realizing pj |(acl() {g }) such that |= i (g 1 b). Let a = g 1 b. By 4.4, tp(a/acl()) is generic (and independent from g over ). As |= i (a), tp(a/acl()) = pi . So g a = b, so g pi = pj . The converse is similar. 48

Proposition 4.8 The index of G0 in G is precisely d, the Morley degree of G. Hence G0 and each of its translates has Morley degree 1, and the generic types p1 , .., pd of G correspond to the cosets of G0 in G. Proof. Let r be the index of G0 in G. As G0 and each of its cosets in G have Morley rank , (*) r d. Now consider the action of G on S described above. Let Gi = F ix(pi ) = {g G : g pi = pi }. By transitivity of the action (4.7 (i)), each Gi has index d in G. By 4.7 (ii), each Gi is denable, so contains G0 , and hence by (*), equals G0 . We worked above with a left action of G on S . Identical results hold for the corresponding right action. Lemma 4.9 Let X be a denable subset of G. Then X is generic if and only if nitely many left translates of X cover G if and only if nitely many right translates cover G. Proof. Suppose rst that nitely many left (or right) translates of X cover G. As each such translate has the same Morley rank as X , it follows that RM (X ) = hence X is generic. Now for the converse. Suppose X to be generic. To prove that nitely many left translates of X cover G, we may clearly replace X by a translate of a denable subset of X . Note that some generic denable subset of X is contained in a single coset of G0 in G. So after translating, we may assume that X is contained in G0 . It is clearly enough to show that nitely many translates of X cover G0 . Assume X is A-denable. Let p0 be the unique generic type of G over A containing x G0 . (So p0 is stationary.) Note that x X p0 . (Why?) Let {ai : i < } be a Morley sequence in p0 . Let g G0 . By 2.45, there is ai which is independent from g over A. Then 1 1 1 0 tp(a i g/A) is generic, and ai g G , hence tp(ai g/A) = p0 so contains 1 x X . So a i g X , so g ai X . We have shown that every g G0 is contained in ai X for some i < . By compactness, nitely many of the ai X cover G0 . Remark 4.10 In the same way as we dened generic types, connected components,... of G, we can do the same thing for any subgroup of G which is 49

 . If H is such, we call H connected if H = H 0 . So if H is denable in M connected then H has a unique generic type (over any set of parameters over which H is dened). Denition 4.11 Let H be a denable subgroup of G and X a left translate of H in G (namely a left coset c H of H in G). Assume that both H and X are dened over A. By a generic type of X over A we mean some p(x) S (A) such that p(x) |= x X and RM (p) = RM (X )(= RM (H )). Likewise, a X is said to be generic in X over A if tp(a/A) is a generic type of X . (Nte that if H is connected then X has a unique generic type over A.) Lemma 4.12 Let H , X , A be as in Denition 4.11, and let a X . Then a is generic in X over A if there is b X such that a is independent from b over A and b1 a is generic in H over A {b}. Proof. Suppose RM (X ) = RM (H ) = . First suppose a to be generic in X over A. Let b be any element of X such that a is indepependent from b over A. So RM (tp(a/A {b})) = . So RM (tp(b1 a/A {b})) = . But b1 a H , so is generic in H over A {b}). The converse is similar. Let us now discuss stabilizers which were already implicit in the material above. Denition 4.13 Let p(x) = tp(a/A) be stationary, where a G. By the left stabilizer of p(x) we mean {g G : for some a realizing p(x)|(A {g }), tp(g a /A) = p(x)}. We write Stab(p(x)) for this left stabilizer. Remark 4.14 Suppose p(x) = tp(a/A) is stationary, with a G. (i) Let g G and a realize p(x)|(A {g }). Then tp(g a/A) = p(x) if and only if tp(g a/A {g }) = p(x)|(A {g }). (ii) Let (x) be a formula in p(x) of Morley rank equal to RM (p) and of Morley degree 1. Then g Stab(p) i g ((G)) (G) has Morley rank . (iii) The left stabilizer of p(x) is equal to the right stabilizer of p1 (x) = tp(a1 /A). Proof. (i) Right to left is clear. Suppose now that tp(g a /A) = p(x). Clearly RM (tp(g a /A {g })) = RM (tp(a /A {g })) = RM (p(x)). So g a is independent from g over A. (ii) and (iii) are left to the reader. 50

Lemma 4.15 Let p(x) S (A) be stationary and contain x G. Then (i) Stab(p) is an A-denable subgroup of G. (ii) if p (x) S (B ) is a nonforking extension of p(x), then Stab(p ) = Stab(p), (iii) RM (Stab(p)) RM (p). Proof. (i) Let (x) p(x) have Morley rank = RM (p) and Morley degree 1. Let g G. As in the proof of 4.7(ii), we see that g Stab(p) i (g 1 x) is contained in the global nonforking extension of p(x), and that this is an Adenable condition on g . By Remark 4.14(i), g Stab(p) i g 1 Stab(p). So to see that Stab(p) is a subgroup it suces to see that it is closed under multiplication. Suppose g, h Stab(p). Let a realize p independent of h g over A. We may assume a is independent from {g, h} over A. So g a realises p(x). As in the proof of 4.7(i), g a is independent from h over A. Hence h (g a) realizes p. (ii) By the proof of (i), Stab(p) depends only on the global nonforking extension p of p. (iii) Let g Stab(p). Let a realize p(x)|(A {g }). Then RM (tp(g/A)) = RM (tp(g/A {a})) = RM (tp(g a/A {a})). But the latter is RM (p(x)) as tp(g a/A)) = p(x). Lemma 4.16 Let a G and p = tp(a/A) be stationary Let H = Stab(p). The following are equivalent: (i) RM (H ) = RM (p). (ii) The right coset H a is acl(A)-denable. (iii) H is connected, X = H a is A-denable, and p(x) is the generic type of X over A. Proof. First we know by the previous lemma that H is A-denable. Let X = H a, and let = RM (H ) = RM (X ). (i) implies (ii). Let g H be generic over A {a}. So RM (tp(g/A {a})) = RM (tp(g a/A {a})) = . But as H = Stab(p), tp(g a/A) = p and so has Morley rank . It follows that (*) g a is independent from a over A. But (as g a H a), H is denable over A {a} as well as over A {g a}. Thus by (*) H is denable over acl(A). (Why?) (ii) implies (i). Assuming (ii) we see that the formula x X is in 51

tp(a/acl(A)), and thus RM (X ) RM (tp(a/acl(A)) = RM (p). By the previous lemma we get equality. (ii) implies (iii). Assume X to be acl(A)-denable. To show that X is Adenable we must show that X is invariant under A-elementary permutations of acl(A). Let f be such. As p(x) = tp(a/A) is stationary, tp(a/acl(A)) = tp(f (a)/acl(A)). But a X , and so a f (X ). But f (X ) is clearly also a right translate of H , so must equal X . Clearly p(x) is a generic type of X (we know by (ii) implies (i) that RM (p) = RM (X )). So all that is left is to show that H is connected. Suppose not, and we will contradict the stationarity of p(x). Then H 0 is a proper subgroup of H , also dened over A. So we can nd h1 , h2 H generic over A, with h1 H 0 , and h2 / H 0 , and moreover with hi independent from a over A for i = 1, 2. Then h1 a and h2 a both realize p(x) (as H = Stab(p)), and moreover a rank computation shows that hi a is independent from a over A for i = 1, 2. So tp(h1 a/A {a}) = tp(h2 a/A {a}) = p|(A {a}). But the formula xa1 H 0 is in the rst type but not the second, a contradiction. So H is connected as claimed, and the proof is complete. The theory above, of generics, stabilizers,.. in groups denable in a model of a t.t theory extends easily to denable homogeneous spaces. By an Adenable homogeneous space we mean an A-denable group G together with an A-denable transitive action on an A-denable set X . Note that after naming a point a X , the action of G on X becomes isomorphic to the action of G on G/H where H is F ix(a). Let us now bring canonical bases into the picture. Lemma 4.17 Let X be an A-denable left translate of a connected denable subgroup H of G. Let p(x) S (A) be the generic type of X . Let u be a canonical parameter for X . Then u is interdenable with Cb(p). Proof. Left to the reader. Before the next lemma we will introduce a bit more notation. Suppose p(x) S (A) is a stationary type of an element in G and let H = Stab(p). Let ) be the global nonforking extension of p. Let a realize p in p(x) S (M . For d G, let d p = tp(d a /M ). Note some elementary extension of M that (*) H = {d G : d p = p}. 52

Lemma 4.18 Let p(x) S (A) be a stationary type implying x G. Let H be the left stabilizer of p(x). Let d G and a realize p(x)|(A {d}). Then tp(d a/A {d}) is stationary, and the left coset d H is (as an imaginary) interdenable with Cb(tp(d a/A {d})) over A. (Likewise with right stabilizers and cosets in place of left ones.) Proof. First as tp(a/A {d}) is stationary and a and d a are interdenable over A {d}, also tp(d a/A {d}) is stationary. Note that the global nonforking extension of tp(d a/A {d}) is precisely d p where p is the global nonforking extension of p. By the denition , of canonical bases, we must show that for any A-automorphism f of M f (dH ) = dH i f (dp) = dp). But as f xes A pointwise and p is denable ), f (d p) = f (d) f p = f (d) p. over A, f (p) = p. Thus, for f AutA (M Hence f (d p) = d p i f (d) p = d p i (d1 f (d)) p = p i (by (*) above) d1 f (d) H i d H = f (d) H (using the fact that f (H ) = H as H is A-denable). Lemma 4.19 Let a G be such that tp(a/A) is stationary. Let H = Stab(p). Let c G be generic over A {a}. Then H a (as an imaginary) is interdenable with Cb(stp(c/A {a c})) over A {c} Proof. To make notation easier we assume A = . By genericity of tp(c), a is independent from a c (over ). So a1 is independent from a c. Now by 4.14(iii), H is the right stabilizer of p1 . Hence by Lemma 4.18, H (a c) is interdenable with Cb(tp(a1 (a c)/a c)). As H is -denable, H (a c) is clearly interdenable with H a over c. Hence we get the conclusion of the lemma. We can now obtain some consequences for 1-based groups. Proposition 4.20 Suppose G is 1-based (and -denable). Then (i) Any stationary type tp(a/A) (a G) is the generic type of some right (left) translate of a connected denable subgroup of G. (ii) Any connected denable subgroup H of G is denable over acl(). (iii) G0 is commutative. Proof. (i) Let H = Stab(p) which is A-denable. Let c be generic in G over A {a}. By 4.19, H a (as an imaginary) is interdenable with e = Cb(tp(c/A {a c}) over A {c}. By 1-basedness of G, e acl(c). Hence 53

the imaginary H a is in acl(c). So (as c is independent from a over A, so independent from H a over A), H a acl(A). That is H a is, as a set, acl(A)-denable. Now apply Lemma 4.16 (ii) Let H be connected and denable over A say. Let a be generic in H over A. Let g G be generic over A {a}. So a is generic in H over A {g }. Note that X = g H is dened over A {g }, and tp(a/A {g }) and tp(g a/A {g }) have the same Morley rank (= RM (H ) = say) and that g a X . Hence g a is generic in X over A {g }, and note q (x) = tp(g a/A {g }) is stationary. By Lemma 4.17, (the canonical parameter of) X is interdenable with Cb(q ). By 1-basedness of G, Cb(q ) acl(g a). So X is dened over acl(g a). It then follows that H is dened over acl(g a). By generic choice of g , g a is independent from A over , hence (as H is A-denable) H is dened over acl(). (iii) It is enough to assume that G is connected and prove that G is commutative. Note that the group G G is also 1-based. For g G, let Hg = {(a, ag ) : a G}. Then each Hg is a connected denable subgroup of G G. By (ii), each Hg is acl()-denable. It follows easily that there are only nitely many groups among the Hg . However the set of groups Hg is in bijection with G/Z (G) (via Hg = Hh i g 1 h Z (G)). Hence Z (G) has nite index in G. As G is connected this means that Z (G) = G and so G is commutative. Corollary 4.21 The following are equivalent (where remember that G is a of the t.t theory T ): -denable group in the saturated model M (i) G is 1-based. (iii) For every n < , every denable subset of Gn = G G .. G is a nite Boolean combination of left (right) translate of acl()-denable subgroups of Gn . Proof. (i) implies (ii): Note that as G is 1-based, so is the group Gn . We rst prove: containing x Gn and Claim. Let p and q be complete types over M containing the same right cosets of acl()-denable subgroups of Gn . Then p = q. Proof of claim. By 4.20 there are connected acl()-denable subgroups H1 , H2 of Gn and right translates X1 , X2 of H1 , H2 respectively, such that p 54

 , and q is THE generic type of X2 over is THE generic type of X1 over M M . By the assumptions of the claim, x X1 x X2 is in both p and q. But RM (p) = RM (X1 ) = RM (H1 ) and RM (q) = RM (X2 ) = RM (H2 ). It follows that RM (X1 X2 ) = RM (X1 ) = RM (X2 ) = , say. So p has Morley rank and contains x X2 . That is p is also the generic type of X2 , whence p = q. The claim is proved. over M We can now deduce (ii) from the claim together with a standard compactness argument as in the proof of 2.25 from [1]. Remark 4.22 (ii) in Corollary 4.21 can be replaced by: for any n, any -denable subset of Gn is a nite Boolean combination of translates of denable subgroups. Exercise 4.23 Let M = (M, , .....) be an arbitrary saturated structure (not necessarily t.t), where is a group operation on M . Suppose that every denable subset of M is a Boolean combination of translates of acl() denable subgroups of M . Then T h(M ) is stable.

We will complete this section with a few scattered but important facts about groups in t.t. theories. The rst is a result of Poizat regarding type-denable subgroups of -stable groups(which was subsequently generalized in a far a reaching way by Hrushovski). T remains a complete t.t theory and M saturated model of T . First: Denition 4.24 Let (x) be a partial type, namely a possibly innite but small set of formulas with fee variable x. Assume is closed under nite conjunctions. Then (RM, dM )() = min{RM ((x)), dM ((x))) : (x) (x)}. Exercise 4.25 Suppose that (x) is a partial type over A and that (RM, dM )() = (, d). Then over acleq (A) there are preciselt d complete types which contain (x) and have Morley rank . Let G be our -denable group. We will say that a subgroup H of G is type-denable if it is dened by (i.e. is the solution set) of a partial type (x). 55

Proposition 4.26 Any type-denable subgroup of G is denable, namely dened by a single formula. Proof. Let H be a subgroup of G dened by the partial type (x). Without loss of generality is closed under nite conjunctions and consists of formulas without parameters. (Just add names for parameters in .) Let (RM, dM )() = (, d). By Exercise 4.25, let p1 , .., pd be the complete types over acleq () which contain and have Morley rank . (So the pi can be considered as the generic types of H .) For g G, let g pi = tp(g a/acleq ()) for some (any) a realizing the nonforking extension of pi over {g } acleq (). As remarked earlier, this is well-dened, by stationarity of pi . Let H1 = {g G : for each i = 1, .., d, g pi = pj for some j = 1, ., d}. (So H1 is the stabilizer of {p1 , .., pd }.) As in the proof of Lemma 4.7, H1 is denable (over acleq ()), and it is rather easy to see that H1 is a subgroup of G. Claim. H1 = H . Proof. First let g H1 . Let a realize p1 with a independent of g . Then g a realizes pj for some j . But both pi and pj contain (x), hence a and g a are in H . Thus g H . Conversely, let g H . Let a realize some pi , independent from g over . Then RM (tp(a/g )) = and also RM (tp(g a/g )) = (why?). Hence by the choice of , RM (tp(g a/)) = , and so tp(g a/acleq ()) = pj for some j = 1, .., d. Thus g H1 . The claim is proved. As H1 is denable, the proposition is also proved. A somewhat stronger result than Proposition 4.26 is true. By a type-denable group (in a saturated model) we mean a pair (G, ) such that is a group operation on G, G is a type-denable set, and the graph of is also typedenable. It is not hard to see that there is sone denable function f whose restriction to G G is precisely the group operation. In any case, the result is that if T is t.t, then any such type-denable group is denable. We will leave this as an exercise. Finally we point out how the generic elements of a denable group G control G. Lemma 4.27 For any a G, there are b, c G, each generic in G over a such that a = b c. 56

Proof. Let = RM (G). Let p S (acleq ()) be a generic type of G. Let b realize the nonforking extension of p over a. Then RM (tp(b1 /a)) = , so RM (tp(b1 a/a) = . Putting c = b1 a, we see that c is generic over a and a = b c. Corollary 4.28 Suppose that X is a -invariant set, and that some generic type of G is X -internal: namely there is a generic type p(x) S (acleq ()) of G, a small set A acleq () of parameters, a realization a of p(x)|A and some tuple c from X such that a dcl(A, c). THEN there is a small set B of eq parameters such that G dcl(B X ) (namely G XB ). Likewise replacing internal by almost internal and dcl by acl. Proof. Let e1 , .., ed be representatives of cosets of G0 in G. Let B = A {e1 , .., ed }. Let a G. Let b, c G be given by Lemma 4.27. Without loss of generality, (b, c) is independent from {a} B over {a}. There are ei and ej such that b = b ei realizes p(x)|B and c = c ej realizes p(x)|B . By assumptions each of b , c is in the denable closure of A together with some tuple from X . It follows that a = b c is in the denable closure of B together with a tuple from X .

Theories and groups of nite Morley rank

The aim of this section is to give some kind of coherent picture of the category of denable sets in a (many-sorted saturated) structure in which all denable sets have nite Morley rank. a saturated model of T . For now T remains a complete t.t. theory and M eq . There is no harm in working in M Denition 5.1 We dene the U -rank on complete types. p(x), q (x), .. denote . complete types over subsets of M (i) U (p) 0 for all p. (ii) U (p) +1 if p(x) has a forking extension q (x) such that U (q ) . (iii) For limit , U (p) if U (p) for all < . (iv) U (p) = if is the least ordinal such that U (p) . (v) U (p) = if U (p) for all ordinals . Lemma 5.2 For any complete type p(x), U (p) RM (p). 57

Proof. We prove by induction on that U (p) implies RM (p) for all cmplete types). The only thing to prove is the induction step: Suppose U (p) +1, so p has a forking extension q (x) with U (q ) . By Proposition 2.26, RM (q ) < RM (p), and by induction hypothesis, RM (q ) + 1. Hence RM (p) . So in our context (T t.t.) every complete type has ordinal-valued U -rank. Note that if p(x) q (x) then U (p) U (q ) (from the denitions). Note also then U (p) = U (f (p)). that if if p S (A) and f is an automorphism of M Also it is rather easy to see that U (p) =) i p is algebraic (has only nitely many realizations). We will need the following forking exercise. Exercise 5.3 Suppose that A B and A C . Let p(x) S (A), q (x) S (B ) and r(x) S (C ) be such that q (x) is a nonforking extension of p(x) and r(x) is an extension of p(x). Then there is an A-automorphism f of M and a complete type q (x) S (B f (C )) such that q (x) q (x) and q (x) is a nonforking extension of f (r) S (f (C )). Lemma 5.4 For any complete types p(x) q (x), U (p) = U (q ) if and only if q is a nonforking extension of p. Proof. Left to right is immediate from the denition. Now suppose that p(x) S (A), A B and q (x) S (B ) is a nonforking extension of p(x). Suppose that U (p) = . We want to prove that U (q ) = , that is, (*) for any < , q has a forking extension with U -rank . Now p(x) has a forking extension r(x) S (C ) such that U (r) . By the exercise above, we may assume that r(x) has a nonforking extension q (x) S (B C ) which extends q (x). Note that U (r) = for some < So by induction hypothesis, U (q ) = . On the other hand, clearly q is a forking extension of q (why?). We have proved (*), and hence the lemma. Here is a nice additivity property for types of nite U -rank, which is proved using Corollary 2.31. Exercise 5.5 Suppose that U (tp(b/A) and U (tp(a/A {b}) are both nite. Then U (tp(a, b/A)) = U (tp(a/A {b})) + U (tp(tp(b/A)).

58

Complete types of U -rank 1 will be important for us. Remark 5.6 Let p(x) S (A) be a stationary type. The following are equivalent: (i) U (p) = 1, (ii) For any B A, p(x) has a unique nonalgebraic extension q (x) S (B ). (iii) p is nonalgebraic, and for any formula (x) (maybe with parameters), the set of realizations of p(x) {(x)} is either nite, or is conite in the set of realizations of p(X ). Types satisfying the equivalent conditions of Remark 5.6 are called minimal types. For example, if (RM (p), dM (p)) = (1, 1) then p is minimal. But there are examples of minimal types with Morley rank > 1: Let the language L consist just of unary predicates Pi for i . Let T say that the Pi are innite and pairwise disjoint. T has quantier-elimination (by a back-andforth argument) from which we can also conclude that each Pi is strongly minimal, and the universe (x = x) has Morley rank 2 and Morley degree 1. Let p(x) S () be the generic type of x = x, namely p(x) is the unique 1-type over with Morley rank 2. p(x) is (axiomatized by) {Pi (x) : i < }. p is stationary (as it has Morley degree 1). We claim that p is minimal: Let B be any set of parameters, and q (x) S (B ) a nonalgebraic extension of p(x). By quantier-elimination we see that q (x) is axiomatized by p(x) {x = b : b B }. So q (x) is the unique nonalgebraic extension of p(x) over B . Exercise 5.7 Let p(x) S (A) be a minimal type. Let X be the set of . Then X together with algebraic closure over A, realizations of p(x) in M is a homogeneous pregeometry in the sense of section 3. Moreover, if a is a nite sequence of elements of X , then dim(a) = U (tp(a/A)). All the results of section 3 on strongly minimal sets and their pregeometries are valid for X (with U -rank replacing Morley rank). On the other hand we have: Lemma 5.8 Suppose that p S (A) is minimal, and that the corresponding pregeometry is nontrivial. Then p(x) has Morley rank 1 (so is the generic type of a strongly minimal set over A). 59

Proof. It is enough to nd some nonforking extension of p(x) which has Morley rank 1 (by 2.26). Thus, by the denition of nontriviality, we may nd realizations a, b, c of p(x) which are pairwise independent over A but such that each of a, b, c is in the algebraic closure of A together with the other two. Let (x, y, z ) be a formula over A witnessing this (that is |= (a, b, c) and |= x, y k z ((x, y, z )) etc.) We may also assume that (x, b, c) isolates tp(a/A, b, c). Now suppose that RM (p) = and let (x) p(x) be chosen of Morley rank and Morley degree 1. (So for any B A, the nonforking extension of p over B is the unique type over B which contains and has Morley rank . By denability of the stationary type p(x), we may nd a formula (x) over A, such that for any a , (a ) i for b realizing p(x)|(A, a ), z ((a , b , z ) (x)). Note that |= (a), hence (x) p(x) and (x) (x) p(x). Let (x) = (x) (x). Then (x) has Morley rank and Morley degree 1. Claim. For any B A, there is a unique complete nonalgebraic type over B which contains (x) (so has to be the unique nonforking extension of p(x) over B ). Proof of claim. Let B A. Let a realize (x) with a / acl(B ). Let b realize p(x)|(B, a ). By choice of , there is c such that |= (a , b , c , ) (c ). Now by choice of , b acl(A, a , c ), hence RM (tp(b /B, a )) RM (tp(c /B, a )) RM (tp(c /B ) (as |= (c ). On the other hand, RM (tp(b /B, a )) = . Hence we conclude that: (*) RM (tp(c /B ) = , and so c realizes p(x)|B . Now if c acl(B, b ) then (by choice of ), a acl(B, b ), but a is independent from b over B , so a acl(B ), a contradiction. So c / acl(B, b ), so by (*) realizes p(x)|(B, b ). Forking calculus implies newline (**) (b , c ) is independent from B over A, and also (***) tp(b , c /A) = tp(b, c/A). As |= (a , b , c ), the choice of together with (***) implies that tp(a /A) = p(x). Also by (**) and the fact that a acl(A, b , c ) we see that a is independent from B over A. Hence a realizes p(x)|B . The claim is proved. It follows from the claim that (x) is strongly minimal. (If (x) were the disjoint union of two innite denable sets, then we obtain two nonalgebraic complete types containing (x) over some set of parameters.)

60

Types of U -rank 1 play an important role in that they coordinatize denable sets of nite Morley rank, in a sense we will now make precise. Proposition 5.9 Let p(x) S (A) be nonalgebraic and of nite U -rank. Then p(x) is nonorthogonal to some stationary type p0 of U -rank 1 (in the sense of Denition 3.36). (Moreover if RM (p) < then RM (p0 ) < .) Proof. Suppose U (p) = n and we may suppose that n > 1. By denition of U -rank, there is B A and an extension q (x) S (B ) of p(x) such that U (q ) = n 1. We may assume that B = acleq (B ) and so q is stationary. eq be the canonical base of q (or rather an element such that Let c M Cb(q ) dcleq (c) as in Lemma 2.38). By 2.42, c is in the denable closure of a nite sequence of realizations of p. By 5.5, U (tp(c/A)) is nite (and clearly nonzero). Let a realize q and note that U (tp(a/A, c)) = n 1. Now, we can again nd some set D of parameters containing A such that U (tp(c/D)) = 1. Without loss D is independent from a over A {c} (why?). Claim. a is independent from D over A. Proof of Claim. By choice of D and forking calculus, tp(a/Dc) does not fork over Ac, and so has U -rank n 1. But U (a/D) U (a/Dc), so if a forked with D over A, we would have to have that U (a/D) = n 1, and so tp(a/Dc) is a nonforking extension of tp(a/D). But c is the canonical base of the latter type, so by 2.38 (ii), we would deduce that c acl(D), a contradiction to U (tp(c/D)) = 1. The claim is proved. Note that a forks with c over D. (Otherwise, by transitivity of nonforking and the claim, it would follow that a is independent from Ac over A.) Thus (by the claim), tp(a/A) = p is nonorthogonal to tp(c/D), proving the proposition. Note that c acl(D, a) so RM (tp(c/D)) RM (tp(a/D)). Remark 5.10 A similar kind of argument, together with a suitable generalization of 5.5 to the innite U -rank case, yields that any stationary type is nonorthogonal to a type whose U -rank is of the form for some ordinal . Denition 5.11 Let p(x) S (A) be stationary. Then p is said to be semiminimal if there are B A, a realising p|B and c1 , .., cn each of whose type over B is minimal, such that a dcl(B, c1 , .., cn ). Proposition 5.12 Suppose that a / acl(A), and U (tp(a/A)) < . Then there is d dcl(a, A) \ acl(A) such that stp(d/A) is semi-minimal. (Again 61

if RM (tp(a/A)) < then stp(d/A) is semi-minimal with respect to minimal types of nite Morley rank.) Proof. By 5.9 there are B A, such that a is independent from B over A, and c such that tp(c/B ) is minimal, and a forks with c over B (so c acl(B, a)). eq (why?). Let We may assume that B = A {b} for some element b M d = Cb(stp(b, c/A, a)). By 2.42, there is a Morley sequence (bi , ci )i< such that d dcl(b0 , c0 , .., bn , cn ) for some n. Let B = acl(A {b0 , .., bn }. Forking calculus yields that a is independent from B over A. Thus, as d acl(A, a), d is independent from B over A. For each i, tp(ci /A, bi ) is minimal, and either tp(ci /B ) is the unique nnforking extension of the latter type, or ci B . Thus, d is contained in the denable closure of B together with some realizations of minimal types over B . That is, tp(d/A) is semi-minimal. Now we only know that d acl(A, a). Let d be the imaginary {d1 , .., dm } where the dj are the conjugates of d over A, a. Then d dcl(A, a) and we leave it to the reader to check that tp(d /A) is still semi-minimal. Let us see what 5.12 means at the level of denable sets. Suppose that X is a denable set of nite Morley rank and of Morley degree 1, denable over A say. Let a realize the generic type of X over A. So there is an A-denable function f such that f is dened on a and stp(f (a)/A) is semi-minimal. In fact, as tp(a/A) is stationary, so is tp(f (a)/A). So f (a) realizes the generic type of some A-denable set Z of nite Morley rank and Morley degree 1. By semiminimality and compactness, after rening Z , there are B A, B denable sets Y1 , .., Yk each of nite Morley rank and Morley degree 1, whose generic types are minimal, and such that Z (Y1 Y2 ..Yk )eq B . Again by compactness, after throwing away from X an A-denable set of small Morley rank, f (X ) Z . So putting it altogether, there is an A-denable function f from X onto an A-denable set Z which lives inside a product of denable sets whose generic types are minimal. For b Z over A, we have the bre f 1 (b), dened over A, b. We have an A, b-denable equivalence relation on f 1 (b) whose classes have Morley degree 1 and we can again apply 5.9 to these classes. By compactness the resulting maps are uniformly denable (as b varies in Z ). The process must eventually stop. So we have a bration of X by denable sets contained in products of denable sets with minimal generic types. If it so happens that all minimal types are nontrivial, then by 5.8, 62

the denable sets whose generic types are minimal are precisely the strongly minimal sets.

Let us now pass to the ne structure of groups of nite Morley rank. is saturated model of T , and G Assumption. T is a complete t.t. theory, M is a group denable in M with RM (G) < . We want to explain how G is built up from almost strongly minimal(or semi-minimal) groups, and then discuss the socle (greatest connected semiminimal subgroup) of G. There are two approaches, from below, using the so-called Zilber indecomposability theore, or from above, using 5.12. We will discuss both. Denition 5.13 Let X be a type-denable subset of G (that is X is the set of realizations of some partial type (x)). We say that X is indecomposable if for each denable subgroup H of G, either X is contained in a single left coset a H , or innitely many left cosets of H meet X . Note that if X is indecomposable, then so is a X for any a G. In any case here are some examples: Lemma 5.14 Suppose that X is the set of realizations of a complete stationary type p(x) S (A) (where G is dened over A, and p(x) |= x G). Then X is indecomposable. Proof. We may assume that A = . Let RM (p) = m. Suppose by way of contradiction, that H is a denable subgroup of G and |X/H | is nite with cardinality > 1. It follows that there is a unique coset, say a H such that RM (X a H ) = m. (Let B be such that H and each coset of H meeting X are dened over B . Then the nonforking extension of p(x) over B must contain exactly one of the x ai H where a1 H, .., ar H are the cosets meeting X .) Now, using the DCC on denable subgroups of G, let K < H be the (unique) smallest denable subgroup of G such that for some coset b K of K in G, RM (X b K ) = m. Note that b K a H . Claim. b K is -denable (and so K is too). . Then f (b K ) = f (b) f (K ) Proof of Claim. Let f be an automorphism of M and RM (X f (b) f (K )) = m. It follows that RM (X b K f (b) f (K )) = m 63

(why?). Hence by choice of K , and stationarity of p, f (b K ) = b K . This proves the claim. But p is the set of realizations of a complete type over , so by the claim p(x) |= x b K , hence X is contained in b K and so also in a H , contradiction. The proof of the lemma is complete. Note that we actually proved above that if p(x) is a stationary type, and p(x) |= x G and some nonforking extension of p contains x a H for some denable subgroup H of G, then all realizations of p are contained in a H. Lemma 5.15 Suppose that G is connected. Then for any a G, the conjugacy class aG is indecomposable. Proof. Suppose again for a contradiction that H is a denable subgroup of G and aG /H has cardinality nite and > 1. For each g G, conjugation by g is a (group) automorphism of G which leaves aG invariant setwise, and takes H to the denable subgroup H g . Hence the same is true for aG /H g for all g G. By the DCC on denable groups, the intersection of all conjugates H g of H is a nite subintersection, so if K is this intersection, then K is normal in G and again aG /K is nite of cardinality > 1. So we see that working in the group G/K , the conjugacy class (a/K )G/K is nite and of cardinality > 1. But this conjugacy class is in denable bijection with ((G/K )/(CG/K (a/K ))), the centralizer of a/K in G/K . Thus this centralizer has nite index > 1 in G/K . But G/K is connected (as G is), and we have a contradiction. The following is proved using the DCC on denable groups. Exercise 5.16 Any denable subset X of G can be partitioned into a nite number of denable indecomposable sets. Proposition 5.17 Suppose that {Xi : i I } is a set of indecomposable type-denable subsets of G such that the identity element e G is in each Xi . Then the (abstract) subgroup of G generated by the Xi s is denable and connected. Moreover H = Xi1 .. Xin for some i1 , .., in I .

64

Proof. Again we may assume G to be dened over . For i1 , .., in I , let Xi1 ... Xin = {a1 ... an : aj Xij }, a type-denable subset of G. As G has nite Morley rank, we can nd such Y = Xi1 ... Xin (for some ii , .., in I ) such that RM (Y ) = m is maximized. Suppose Y is type-denable over A = acl(A), and let p(x) S (A) be (stationary) and of Morley rank m such that p(x) |= x Y . Let H be the (left) stabilizer of p(x). Claim. Xj H for all j I . Proof of claim. If not, then by indecomposability of Xi , and as e Xj , it follows that Xj /H is innite. Let b1 , b2 , .... be elements in Xj such that br H = bs H for r = s. Let B A be algebraically closed such that Xi is dened over B and each of b1 , b2 , ... is in B . Let p (x) S (B ) be the nonforking extension of p(x) over B , realized by c say. Then for each 1 r < , tp(br c/B ) has Morley rank m. But for each r = s, b / H, s br 1 hence tp(bs br c/B ) = tp(c/B ), hence tp(br c/B ) = tp(bs c/B ). But maximal choice of RM (Y ), RM (Xj Y ) = m. As each br c Xj Y , we see that Xj Y contains innitely many distinct types of Morley rank m, a contradiction. This proves the claim. By the claim, (*) H contains the subgroup of G generated by all the Xi . In particular Y H , and so RM (p) RM (H ). By 4.15 and 4.16, H is connected, and p(x) is the generic type of a translate of H . But by (*), p(x) |= x H . Hence p(x) is the generic type of H . By Lemma 4.27, every element of H is a product of two realizations of p and thus H Y Y . Together with (*) this proves the proposition. I will now give a series of consequences of Proposition 5.17. Remember that of the t.t theory G is a group of nite Morley rank denable in the model M T. Corollary 5.18 Suppose that G is innite, noncommutative and has no innite denable normal proper subgroups. Then (i) G has no innite normal proper subgroup, (ii) Z (G) is nite and G/Z (G) is simple as an abstract group. Proof. (i) Clearly G is connected. By our assumptions, Z (G) (the centre of G) is nite. Suppose H is an innite normal subgroup of G. Let a H \ Z (G). Then aG is innite (otherwise C (a) has nite index in G 65

so C (a) = G so a Z (G), contradiction). By 5.15, aG is indecomposable. So aG b1 is indecomposable and contains e, for all b aG . By 5.17, the subgroup of G generated by all the sets aG b1 for b aG , is denable. Let H1 be this subgroup. Then it is normal, innite, and is clearly contained in H . By our assumptions, H1 = G, hence H = G. (ii) We already know that Z (G) is nite (and normal). Claim I. G/Z (G) is noncommutative and connected. Proof. Clearly G/Z (G) is connected (for otherwise the preimage in G of a proper denable subgroup of nite index of G/Z (G) would contradict connectedness of G. If G/Z (G) were commutative, then G (the subgroup of G generated by all commutators a b a1 b1 ) would be contained in Z (G). But G is generated by all aG a1 for a G, and by 5.15 and 5.17, G is denable, normal (and innite). So G = G, contradiction. Claim II. G/Z (G) has no nite normal subgroups. Proof. The preimage in G of a nite normal subgroup of G/Z (G) would be a nite normal subgroup H of G, containing Z (G). Note that G acts denably on H by conjugation. Hence G/CG (H ) acts faithfully on H , whereby G/CG (H ) is nite. So G = CG (H ) and H = Z (G). Note that G/Z (G) has no proper denable innite normal subgroups. Hence by Claims I, II, and part (i), we see that G/Z (G) has no proper nontrivial normal subgroup. This completes the proof of 5.18. We now want to decompose G into almost strongly minimal groups. Corollary 5.19 Suppose that G is innite and has no proper innite normal denable subgroups. Then G is almost strongly minimal (there is strongly minimal X G and some nite set of parameters B over which G and X eq are dened, such that G XB ). Proof. As RM (G) < , there is some strongly minimal set X G. Let A be such that G and X are A-denable. Let p(x) S (A) be stationary of ). For Morley rank 1 containing X . Let a realize p and let Y = a1 p(M 1 each b G, let Yb = b Y b . Then by 5.14 each Yb is type-denable, indecomposable, and contains e. Let H be the subgroup of G generated by all the Yb s. H is clearly normal, and innite. By part (i) (or 5.17 and the hypotheses), H = G. By 5.17, G = Yb1 ... Ybn for some b1 , .., bn G. Then clearly, for B = A {a, b1 , .., bn }, G dcl(X B ). 66

Now x G, connected and of nite Morley rank. Then there is greatest n such that there exist {e} = H0 < H1 < .. < Hn = G, where each Hi is a normal denable connected subgroup of G (and Hi is a proper subgroup of Hi+1 ). In fact by Lemma 4.1, RM (Hi ) < RM (Hi+1 ), so n RM (G). Note that for each i For future purposes we denote n by n(G). Corollary 5.20 With G, n and the Hi as above, each Hi+1 /Hi is almost strongly minimal. Proof. This is almost given by Corollary 5.19. As there, let X be a strongly minimal subset of Hi+1 /Hi , p(x) S (A) its generic type and Y some by a1 for a realizing p. Now consider the subgroup H/Hi translate of p(M of G/Hi generated by all Y b for b G/Hi . So by 5.17, H/Hi is denable, innite, normal (in G/Hi ), connected, and contained in Hi+1 /Hi . By choice of the Hi , we see that H = Hi+1 . Exercise 5.21 (G a denable group of nite Morley rank.) There are sta eq ) such that tionary types p1 , .., pn of Morley rank 1 (maybe in M (i) each stationary tp(a/A) wth a G is nonorthogonal to one of the p1 , .., pn , (ii) for any stationary type p of U -rank 1, if p is nonorthogonal to tp(a/A) for some a G, then p is nonorthogonal to one of p1 , .., pn (and hence RM (p) = 1 too). Example 5.22 (i) Let T be strongly minimal (e.g. T = T h(ACFp ). Then is almost strongly minimal. any group denable in M ( ) (ii) Let T = T h(Z4 , +). Let G denote this group. Them G is not almost strongly minimal. However both 2G and G/2G) are strongly minimal. We want to say a little more about the Hi+1 /Hi s. Lemma 5.23 Let G be connnected. Let H be a minimal innite normal denable subgroup of G. Then either H is commutative, or H is noncommutative with no proper innite normal subgroup. Proof. Note that H is connected. We proceed by induction on RM (G). We may assume that H = G (using also 5.18). Let us also assume that H is noncommutative. So Z (H ) is nite (otherwise the connected component of 67

Z (H ) is a proper innite normal subgroup of H , normal in G). As in the proof of 5.18, H/Z (H ) is centreless. So (quotienting G by the nite normal subgroup Z (H )) we may assume that H is centreless. It is enough (by 5.18) to show that H has no innite proper normal denable subgroups. Suppose for a contradiction that N is such. So N is connected. By 5.17 the subgroup of G generated by {N g : g G} is denable and connected. This group is clearly innite, normal in G and contained in H . So it must equal H . Thus (by 5.17 again) H = N1 N2 .. Nk for distinct conjugates Ni of N (and note each Ni is normal in H . Now each Ni Nj is normal in H , so nite, so central in H , so trivial. Thus H is a direct product N1 .. Nk . If N were commutative, then each Ni would be too, so H would be commutative, cntradiction. Thus each N i (a minimal normal denable innite subgroup of H ) is noncommutative. We can apply induction (RM (H ) < RM (G)) to conclude that each Ni has no proper innite normal subgroup. Note that Z (Ni ) < Z (H ) = {e}, so each Ni is also centreless. Thus H is a direct product of nitely many noncommutative (abstractly) simple denable groups, N1 , .., Nk . An elementary grou-theoretic fact implies that any simple nontrivial normal subgroup of H is among the Ni . Thus G acts (by conjugation) on the nite set {N1 , .., Nk }. As G is connected, the action is trivial. That is each Ni is normal in G, as is N . Contradiction. The lemma is proved. Corollary 5.24 In the context of 5.20, each Hi+1 /Hi is either commutative, or noncommutative with no innite normal subgroup (so by 5.18 abstractly simple modulo its nite centre). Proof. Hi+1 /Hi is clearly a minimal innite normal subgroup of G/Hi so we can use 5.23. Corollary 5.18 raises the question: if G is connected and innite, does it follow that G has no abstract proper subgroup of nite index. This is not true in general: the additive group of an algebraically closed eld of characteristic p > 0 is connected, but (as it is an innite dimensional vector space over Fp ) it has proper subgroups of nite index. This is essntially the only obstacle. Corollary 5.25 Suppose that for every connected commutative group G of ), G has no proper subgroup of nite index. nite Morley rank (denable in M 68

Then the same is true for every connected denable group of nite Morley rank. Proof. Let n = n(G) and let Hn1 be as in the discussion before 5.20. Suppose for a contradiction that H is a proper subgroup of nite index in G. So H Hn1 has nite index in Hn1 . By induction (RM (Hn1 ) < RM (G)), G contains Hn1 whereby H/Hn1 is a proper subgroup of nite index in G/Hn1 . Note that any proper subgroup of nite index contains a proper normal subgroup of nite index. So we have a contradiction to 5.24, using our assumptions. In particular if all commutative connected denable groups of nite Morley are divisible, then any connected denable group of nite Morley rank is abstractly connected. Finally we discuss the socle argument. G remains a group of nite Morley |= T . rank denable in M Denition 5.26 The socle of G, s(G) is the subgroup of G generated by all the connected denable almost strongly minimal subgroups of G. Remark 5.27 s(G) is connected, denable and normal in G and denable over the same set of parameters as G. Proof. If H is a connected denable subgroup of G, then clearly H is indecomposable and contains the identity. If also H is almost strongly minimal, then every conjugate H g of H is almost strongly minimal (as well as being connected). Hence s(G) is normal. By 5.17 it is denable and connected, and equals K1 .. Kn for some almost strongly minimal denable connected subgroups K1 , .., Kn of G. Lemma 5.28 s(G) is the unique maximal connected denable subgroup H eq with H of G such that there are strongly minimal sets Y1 , .., Yd in M acl(Y1 , .., Yd , B ) for some small set B of parameters. Proof. Suppose for a contradiction that H is a connected denable subgroup of G properly containing S = s(G) such that H is contained in the algebraic closure of a nite number Y1 , .., Yd of strongly minimal sets (plus a small 69

set of parameters). Let B be the set of parameters (over which we may assume that G, S, H and the Yi are dened). Let a H with tp(a/B ) generic. Then a/S / acl(B ). Let bi Yi for i = 1, .., d,such that a acl(B, b1 , .., bd ). We may assume that for each i bi / acl(B {bj : j = i}). Note that a/S acl(B, b1 , .., bd ) too. After including in B some of the bi s, we may assume that each of a, a/S is interalgebraic with (b1 , .., br ) over B and that a/S / acl(B ). (We still have that (b1 , .., br ) is A-independent, but tp(a/B ) need no longer be generic in H .) In particular, both a/S and a are interalgebraic with br over B, b1 , .., br1 . So both stp((a/S )/B, b1 , .., br1 ) and q (x) = stp(a/B, b1 , .., br1 ) are stationary of Morley rank 1. In particular the set of realizations X of q (x) intersects innitely many cosets of S in G. The same is true for a1 X . By 5.17 the latter set generates a connected denable almost strongly minimal subgroup of G which is not contained in S , contradiction. We will restrict our attention to commutative groups below, although everything works in the noncommutative case too. Assume G to be dened over . Proposition 5.29 Let G be connected and commutative (and of nite Morley rank). Let S = s(G). Let p(x) = tp(a/A) be stationary, with a G. Assume also (i) every connected denable subgroup of S is dened over some c such that c is independent from d over A for all d G/S . (ii) Stab(p) is nite. Then all realizations of p(x) are contained in a single translate of S . Before proving the proposition, let is discuss the hypotheses and conclusion. Hypothesis (i) is a rigidity condition. It will be satised in the following special situations: (a) every denable connected subgroup of G is dened over acl(), (b) S is fully orthogonal to G/S (whenever p(x) and q (y ) are complete stationary types extending x S , y G/S respectively, then p is orthogonal to q ). (a) will be true of algebraic tori (nite products of the is multiplicative group) as well as abelian varieties, in the case where M an algebraically closed eld, although in this case every denable connected group is equal to its socle. A more interesting example is where G as a structure in its own right is a compact complex Lie group equipped with 70

predicates for analytic subvarieties of Gn for all n. (Such a structure is not saturated but we can work with a saturated elementary extension). Hypothesis (ii) is equivalent to Stab(p) S is nite. For suppose the latter holds and Stab(p) is innite. Then the connected component of Stab(p) contains an almost strongly minimal subgroup (why?), which is not contained in S , a contradiction. The conclusion of the Proposition could possibly be better expressed in terms of denable sets. For X G a denable set of Morley rank m and Morley degree 1, dene Stab(X ) to be {g G : RM (X g X ) = m}. Then Stab(X ) is the same as Stab(p) where p is the unique stationary type of Morley rank m extending x X . Then the conclusion can be stated as: if Stab(X ) is nite, then, up to translation, and a denable set of Morley rank < m, X is contained in S . Together with the truth of the Zilber conjecture in the many sorted structure of compact complex manifolds (which will be discused later), the Proposition yields the following result of Ueno: if X is an irredicible analytic subvariety of a compact complex Lie group A, and Stab(X ) is nite, then X is biholomorphic with a complex algebraic variety. (However there are simpler proofs, even of a model-theoretic nature.) Proof of Proposition 5.29. The proof is a little bit involved. The general strategy is to assume the conclusion is false, and then produce a connected denable nontrivial subgroup of G which is contained in the algebraic closure ) and is NOT contained in S , of nitely many strongly minimal sets (in M contradicting 5.28. We may assume that A = . Let : G G/S be the canonical surjective homomorphism. Let X be the set of realizations of p. Fix a X and let b = (a). Note that tp(b/) is also stationary, so if it is algebraic then b dcl(), and we see that all realizations of p are contained in the single translate 1 (b) of S . So we may assume that b / acl(). Let Xb = X 1 (b). Note that Xb is precisely the set of realizations of tp(a/b) (why?). Let Gb = 1 (b). So Gb is a translate of S . In particular S acts on Gb by addition. Now let Z be a subset of Gb which is denable with parameters from S {b} and is of of least (RM, dM ) subject to this condition. Claim I. Z is of the form H + c for some c Z and some denable subgroup H of S . Proof of Claim I. Note that if s S , then either s + Z = Z or s + Z is 71

disjoint from Z . (If both (s + Z ) Z and the symmetric dierence of s + Z and Z were nonempty then one would have strictly smaller (RM, dM ) than Z , a contradiction, as both are denable over S {b}.) Thus H = {s S : s + Z = Z } is a denable subgroup of S , and Z = H + c for some (any) c Z. Note that the sets s + Z for s S cover Gb and each is a coset (translate) of H. Claim II. Any translate s + Z of Z by s S is either contained in Xb or is disjoint from Xb . Thus Xb is invariant under translation by H . Proof. Suppose that s + Z meets Xb . Now Xb is an intersection of b-denable sets {Yi }i say, so as s + Z is dened over S {b} and has the same Morley rank and degree as Z , s + Z must be contained in every Yi (otherwise we again contradict choice of Z ). So s + Z is contained in Xb . Now let H 0 be the connected component of H (a connected denable subgroup of S ). By Claim II, Xb is invariant under translation by H 0 . Claim III. H 0 is contained in Stab(p). Proof. By assumption, H 0 is dened over some parameter t such that t is independent from b over . Fix d H 0 . We will show that d Stab(p). We may assume that d is independent from b over t, hence {t, d} is independent from b over . Let a Xb be independent from {t, d} over b. Then by forking calculus we see that a is independent from d over (and a realizes p). As Xb is invariant under translation by H 0 , d + a Xb so realizes p too. This proves that H0 Stab(p). From Claim III and assumption (ii), H 0 is trivial hence H is nite, so Z is nite. Z is dened by a formula (x, b, e) for some e from S . We may assume that for all b , e , (x, b , e ) denes a nite (maybe empty) set. As tp(b) is nonalgebraic, there is an innite denable subset Y of G/S such that for all b Y , there is e from S such that (x, b , e ) denes a (nite) nonempty subset of the bre Gb . As Y is innite, it contains a strongly minimal denable subset Y say. Let W be the the union of all sets dened by (x, b , e ) for b Y and e in S . Thus W is a denable subset of G contained in the algebraic closure of S and Y . As S itself is contained in the denable closure of a nite set of strongly minimal sets (plus some parameters), W is contained in the algebraic closure of nitely many strongly minimal sets. Note that W meets innitely many cosets of S in G. By Exercise 5.15, there

72

is some indecomposable denable subset W of W which meets innitely many cosets of S in G. After translating W we obtain an indecomposable denable subset W of G which contains the indentity and meets innitely many cosets of S . Let K be the subgroup of G generated by W . By 5.17, K is denable, connected, not contained in S . The (denable connected) subgroup of G generated by S and K is then contained in the algebraic closure of nitely many strongly minimal sets, and properly contains S . This contradicts Lemma 5.28. Proposition 5.29 is proved. It is an open question whether Proposition 5.29 is true without the rigidity assumption (i).

References
[1] A. Pillay, Lecture Notes - Model Theory.

73